Search

In recent years, the construction industry has been striving to make production faster and handle more complex architectural designs. Waste reduction, geometric freedom, lower construction costs, and speedy construction make the 3D-printed fiber-reinforced concrete (3DPFRC) alternative for future construction. However, achieving the optimum mixture composition for 3DPFRC remains a daunting task, entailing the consideration of multiple variables and necessitating an extensive trial-and-error experimental process. Therefore, this study investigated the application of different metaheuristic optimization algorithms to predict the compressive strength (CS) of 3DPFRC. A database of 299 data samples with 16 different input features was compiled from the experimental studies in the literature. Six metaheuristic algorithms, such as human felicity algorithm (HFA), differential evolution algorithm (DEA), nuclear reaction optimization (NRO), Harris hawks optimization (HHO), lightning search algorithm (LSA), and tunicate swarm algorithm (TSA) were applied to identify the optimal hyperparameter combination for the random forest (RF) model in predicting the CS of 3DPFRC. Different statistical metrics and 10-fold cross-validation were used to evaluate the accuracy of the models. The TSA-RF model exhibited superior performance compared to other models, achieving correlation (R), mean absolute error (MAE), and root mean square error (RMSE) values of 0.99, 2.10 MPa, and 3.59 MPa, respectively. The LSA-RF model also performed well, with R, MAE, and RMSE values of 0.99, 2.93 MPa, and 6.23 MPa, respectively. SHapley Additive exPlanation (SHAP) interpretability elucidates the intricate relationships between features and their effects on the CS, thereby offering invaluable insights for the performance-based mix proportion design of 3DPFRC., Validerad;2024;Nivå 2;2024-01-25 (signyg);Funder: Najran University (NU/DRP/SERC/12/10);Full text license: CC BY-NC-ND

Locoregional recurrences and distant metastases are major problems for patients with squamous cell carcinoma of the head and neck (SCCHN). Because SCCHN is a heterogeneous group of tumours with varying characteristics, the present study concentrated on the subgroup of squamous cell carcinoma of the oral tongue (SCCOT) to investigate the use of machine learning approaches to predict the risk of recurrence from routine clinical data available at diagnosis. The approach also identified the most important parameters that identify and classify recurrence risk. A total of 66 patients with SCCOT were included. Clinical data available at diagnosis were analysed using statistical analysis and machine learning approaches. Tumour recurrence was associated with T stage (P=0.001), radiological neck metastasis (P=0.010) and diabetes (P=0.003). A machine learning model based on the random forest algorithm and with attendant explainability was used. Whilst patients with diabetes were overrepresented in the SCCOT cohort, diabetics had lower recur‑ rence rates (P=0.015 after adjusting for age and other clinical features) and an improved 2‑year survival (P=0.025) compared with non‑diabetics. Clinical, radiological and histological data available at diagnosis were used to establish a prognostic model for patients with SCCOT. Using machine learning to predict recurrence produced a classification model with 71.2% accuracy. Notably, one of the findings of the feature importance rankings of the model was that diabetics exhibited less recur‑ rence and improved survival compared with non‑diabetics, even after accounting for the independent prognostic variables of tumour size and patient age at diagnosis. These data imply that the therapeutic manipulation of glucose levels used to treatdiabetes may be useful for patients with SCCOT regardless of their diabetic status. Further studies are warranted to investigatethe impact of diabetes in other SCCHN subtypes.

In recent years, the construction industry has been striving to make production faster and handle more complex architectural designs. Waste reduction, geometric freedom, lower construction costs, and speedy construction make the 3D-printed fiber-reinforced concrete (3DPFRC) alternative for future construction. However, achieving the optimum mixture composition for 3DPFRC remains a daunting task, entailing the consideration of multiple variables and necessitating an extensive trial-and-error experimental process. Therefore, this study investigated the application of different metaheuristic optimization algorithms to predict the compressive strength (CS) of 3DPFRC. A database of 299 data samples with 16 different input features was compiled from the experimental studies in the literature. Six metaheuristic algorithms, such as human felicity algorithm (HFA), differential evolution algorithm (DEA), nuclear reaction optimization (NRO), Harris hawks optimization (HHO), lightning search algorithm (LSA), and tunicate swarm algorithm (TSA) were applied to identify the optimal hyperparameter combination for the random forest (RF) model in predicting the CS of 3DPFRC. Different statistical metrics and 10-fold cross-validation were used to evaluate the accuracy of the models. The TSA-RF model exhibited superior performance compared to other models, achieving correlation (R), mean absolute error (MAE), and root mean square error (RMSE) values of 0.99, 2.10 MPa, and 3.59 MPa, respectively. The LSA-RF model also performed well, with R, MAE, and RMSE values of 0.99, 2.93 MPa, and 6.23 MPa, respectively. SHapley Additive exPlanation (SHAP) interpretability elucidates the intricate relationships between features and their effects on the CS, thereby offering invaluable insights for the performance-based mix proportion design of 3DPFRC., Validerad;2024;Nivå 2;2024-01-25 (signyg);Funder: Najran University (NU/DRP/SERC/12/10);Full text license: CC BY-NC-ND

Locoregional recurrences and distant metastases are major problems for patients with squamous cell carcinoma of the head and neck (SCCHN). Because SCCHN is a heterogeneous group of tumours with varying characteristics, the present study concentrated on the subgroup of squamous cell carcinoma of the oral tongue (SCCOT) to investigate the use of machine learning approaches to predict the risk of recurrence from routine clinical data available at diagnosis. The approach also identified the most important parameters that identify and classify recurrence risk. A total of 66 patients with SCCOT were included. Clinical data available at diagnosis were analysed using statistical analysis and machine learning approaches. Tumour recurrence was associated with T stage (P=0.001), radiological neck metastasis (P=0.010) and diabetes (P=0.003). A machine learning model based on the random forest algorithm and with attendant explainability was used. Whilst patients with diabetes were overrepresented in the SCCOT cohort, diabetics had lower recur‑ rence rates (P=0.015 after adjusting for age and other clinical features) and an improved 2‑year survival (P=0.025) compared with non‑diabetics. Clinical, radiological and histological data available at diagnosis were used to establish a prognostic model for patients with SCCOT. Using machine learning to predict recurrence produced a classification model with 71.2% accuracy. Notably, one of the findings of the feature importance rankings of the model was that diabetics exhibited less recur‑ rence and improved survival compared with non‑diabetics, even after accounting for the independent prognostic variables of tumour size and patient age at diagnosis. These data imply that the therapeutic manipulation of glucose levels used to treatdiabetes may be useful for patients with SCCOT regardless of their diabetic status. Further studies are warranted to investigatethe impact of diabetes in other SCCHN subtypes.

In the era of Industry 4.0 and 5.0, a transformative wave of softwarisation has surged. This shift towards software-centric frameworks has been a cornerstone and has highlighted the need to comprehend software applications. This research introduces a novel agent-based architecture designed to sense and predict software application metrics in industrial scenarios using AI techniques. It comprises interconnected agents that aim to enhance operational insights and decision-making processes. The forecaster component uses a random forest regressor to predict known and aggregated metrics. Further analysis demonstrates overall robust predictive capabilities. Visual representations and an error analysis underscore the forecasting accuracy and limitations. This work establishes a foundational understanding and predictive architecture for software behaviours, charting a course for future advancements in decision-making components within evolving industrial landscapes.

Background. The Android smartphone, with its wide range of uses and excellent performance, has attracted numerous users. Still, this domination of the Android platform also has motivated the attackers to develop malware. The traditional methodology which detects the malware based on the signature is unfit to discover unknown applications. In this paper, detection has been tried whether an application is malware or not using Static Analysis (SA). Considered all the permissions that an application asks for and took them as input to feed our machine learning models. Objectives. The objectives to address and fulfill the aim of this thesis are: To find/create the necessary data set containing malware in the android systems. To test this, different classifiers have been built using different machine learning (ML) algorithms such as Support Vector Machine (SVM) (Linear and RBF), Logistic Regression (LR), Random Forest Algorithm (RF), Gaussian Naive-Bayes (GNB), Decision Tree Method (DT) etc., and also compared their performances. To evaluate and compare each of the chosen models using Accuracy, Precision, F1-Score and Recall methods among the algorithms mentioned in detecting the malware in android with better accuracy in real-time scenarios. Methods. To answer the research question, 1 method has been chosen which is: To identify malware in android system, the Experiment has been used. Results. The Sequential Neural Network (SNN) performed well on the dataset with 98.82 percent than the other Machine Learning (ML) algorithms. So, it is the most fruitful algorithm for the Android malware detection. Random Forest (RF), Decision Tree (DT) are the second-best algorithms on the dataset with 97 percent. Conclusions. Among Logistic Regression, KNN, SVM Linear, SVM RBF, Decision Tree, Random Forest, Gaussian Naive Bayes, and Sequential Neural Network Random Forest is declared as the most efficient algorithm after comparing all the models based on the performance metrics Preci

Flygresor.se, a leading flight comparison platform, uses machine learning to rankflights based on their likelihood of being clicked. The main goal of this project was toimprove this flight sorting to obtain a better user experience. The platform's existingmodel is based on a neural network approach and a limited set of features. The solution involved developing and comparing two machine learning models, Random Forest and XGBoost besides using a set of existing and newly created features. TheXGBoost model demonstrated superior performance by significantly improving theprediction of clicked flights by 4.18% while also achieving a remarkable increase inefficiency by being 125 times faster than the existing model., Flygresor.se, en ledande plattform för jämförelse av flygresor, använder maskininlärning för att ranka flygresor baserat på deras sannolikhet att bli klickade. Huvudmåletmed detta projekt var att förbättra denna flygsortering för att få en bättre användarupplevelse. Plattformens befintliga modell är baserad på ett neuralt nätverk och ettbegränsat antal funktioner. Lösningen innebar att utveckla och jämföra två maskininlärningsmodeller, Random Forest och XGBoost, förutom att använda en uppsättning befintliga och nyskapade funktioner. XGBoost-modellen visade bättre prestandagenom att förbättra predikteringen av de klickade flygresor med 4,18 % samtidigt somden uppnådde högre nivå av effektivitet genom att vara 125 gånger snabbare än denbefintliga modellen.

This study aims to create a Machine Learning model and investigate its performance of detecting fraudulent user behaviour on an e-commerce platform. The user data was analysed to identify and extract critical features distinguishing regular users from fraudulent users. Two different types of user data were used; Event Data and Screen Data, spanning over four weeks. A Principal Component Analysis (PCA) was applied to the Screen Data to reduce its dimensionality. Feature Engineering was conducted on both Event Data and Screen Data. A Random Forest model, a supervised ensemble method, was used for classification. The data was imbalanced due to a significant difference in number of frauds compared to regular users. Therefore, two different balancing methods were used: Oversampling (SMOTE) and changing the Probability Threshold (PT) for the classification model. The best result was achieved with the resampled data where the threshold was set to 0,4. The result of this model was a prediction of 80,88% of actual frauds being predicted as such, while 0,73% of the regular users were falsely predicted as frauds. While this result was promising, questions are raised regarding the validity since there is a possibility that the model was over-fitted on the data set. An indication of this was that the result was significantly less accurate without resampling. However, the overall conclusion from the result was that this study shows an indication that it is possible to distinguish frauds from regular users, with or without resampling. For future research, it would be interesting to see data over a more extended period of time and train the model on real-time data to counter changes in fraudulent behaviour.

Efficient price management is crucial for companies with many different products to keep track of, leading to the common practice of price logging. Today, these prices are often adjusted manually, but setting prices manually can be labor-intensive and prone to human error. This project aims to use machine learning to assist in the pricing of products by estimating the prices to be inserted. Multiple machine learning models have been tested, and an artificial neural network has been implemented for estimating prices effectively. Through additional experimentation, the design of the network was fine-tuned to make it compatible with the project’s needs. The libraries used for implementing and managing the machine learning models are mainly ScikitLearn and TensorFlow. As a result, the trained model has been saved into a file and integrated with an API for accessibility.

Context: Itajai, a coastal city in Santa Catarina, faces unique challenges related to the proliferation of Aedes aegypti, a vector of several diseases. This article presents a predictive model developed to forecast Aedes aegypti hotspots in the region, considering the complex interaction between local climatic variables and global phenomena such as El Niño and La Niña. Methodology: Using a Random Forest algorithm, the model is capable of capturing non-linear relationships in the data, providing insights into the influence of climatic factors on mosquito activity. The choice of this algorithm is due to its robustness and ability to consider the multiplicity of factors influencing mosquito proliferation. By grouping data by season, the model incorporates seasonal nuances, reflecting climatic variations in Itajai. Considerations: The integration of broader climatic patterns highlights the interconnection between local and global factors. This model offers a valuable tool for Itajai's health authorities, enabling proactive actions and resource optimization in the fight against Aedes aegypti. In short, this study proposes an innovative and practical approach to the prevention of mosquito-borne diseases, with potential to positively impact public health in similar regions., Contexto: Itajaí, una ciudad costera en Santa Catarina, enfrenta desafíos únicos relacionados con la proliferación del mosquito Aedes aegypti, vector de varias enfermedades. Este artículo presenta un modelo predictivo desarrollado para prever focos de Aedes aegypti en la región, considerando la compleja interacción entre variables climáticas locales y fenómenos globales como El Niño y La Niña. Metodologia: Utilizando un algoritmo de Random Forest, el modelo es capaz de capturar relaciones no lineales en los datos, ofreciendo perspectivas sobre la influencia de los factores climáticos en la actividad del mosquito. La elección de este algoritmo se debe a su robustez y capacidad para considerar la multiplicidad de factores que influyen en la proliferación del mosquito. Al agrupar datos por estación, el modelo incorpora matices estacionales, reflejando las variaciones climáticas de Itajaí. Resultados: La integración de patrones climáticos más amplios destaca la interconexión entre factores locales y globales. Este modelo ofrece una herramienta valiosa para las autoridades sanitarias de Itajaí, permitiendo acciones proactivas y optimización de recursos en la lucha contra el Aedes aegypti. En resumen, este estudio propone un enfoque innovador y práctico para prevenir enfermedades transmitidas por mosquitos, con el potencial de impactar positivamente en la salud pública en regiones similares., Contexto: Itajaí, uma cidade costeira em Santa Catarina, enfrenta desafios singulares relacionados à proliferação do Aedes aegypti, vetor de várias doenças. Este artigo apresenta um modelo preditivo desenvolvido para prever os focos de Aedes aegypti na região, considerando a interação complexa entre variáveis climáticas locais e fenômenos globais, como El Niño e La Niña. Metodologia: Utilizando um algoritmo Random Forest, o modelo é capaz de capturar relações não-lineares nos dados, oferecendo insights sobre a influência de fatores climáticos na atividade do mosquito. A escolha deste algoritmo se deve à sua robustez e capacidade de considerar a multiplicidade de fatores que influenciam a proliferação do mosquito. Ao agrupar dados por estação do ano, o modelo incorpora nuances sazonais, refletindo as variações climáticas de Itajaí. Considerações: A integração de padrões climáticos mais amplos destaca a interconexão entre fatores locais e globais. Este modelo oferece uma ferramenta valiosa para as autoridades de saúde de Itajaí, permitindo ações proativas e otimização de recursos no combate ao Aedes aegypti. Em suma, este estudo propõe uma abordagem inovadora e prática para a prevenção de doenças transmitidas por mosquitos, com potencial para impactar positivamente a saúde pública em regiões similares.

Annual average land-use regression (LUR) models have been widely used to assess spatial patterns of air pollution exposures. However, they fail to capture diurnal variability in air pollution and consequently might result in biased dynamic exposure assessments. In this study we aimed to model average hourly concentrations for two major pollutants, NO2 and PM2.5, for the Netherlands using the LUR algorithm. We modelled the spatial variation of average hourly concentrations for the years 2016–2019 combined, for two seasons, and for two weekday types. Two modelling approaches were used, supervised linear regression (SLR) and random forest (RF). The potential predictors included population, road, land use, satellite retrievals, and chemical transport model pollution estimates variables with different buffer sizes. We also temporally adjusted hourly concentrations from a 2019 annual model using the hourly monitoring data, to compare its performance with the hourly modelling approach. The results showed that hourly NO2 models performed overall well (5-fold cross validation R2 = 0.50–0.78), while the PM2.5 performed moderately (5-fold cross validation R2 = 0.24–0.62). Both for NO2 and PM2.5 the warm season models performed worse than the cold season ones, and the weekends' worse than weekdays’. The performance of the RF and SLR models was similar for both pollutants. For both SLR and RF, variables with larger buffer sizes representing variation in background concentrations, were selected more often in the weekend models compared to the weekdays, and in the warm season compared to the cold one. Temporal adjustment of annual average models performed overall worse than both modelling approaches (NO2 hourly R2 = 0.35–0.70; PM2.5 hourly R2 = 0.01–0.15). The difference in model performance and selection of variables across hours, seasons, and wee

BACKGROUND: Antimicrobial use (AMU) has decreased significantly in Dutch pig farms since 2009. However, this decrease has stagnated recently, with relatively high AMU levels persisting mainly among weaners. The aim of this study was to identify farm-level characteristics associated with: i) total AMU and ii) use of specific antimicrobial classes.METHODS: In 2020, cross-sectional data from 154 Dutch pig farms were collected, including information on AMU and farm characteristics. A mixed-effects conditional Random Forest analysis was applied to select the subset of features that was best associated with AMU.RESULTS: The main risk factors for total AMU in weaners were vaccination for PRRS in sucklings, being a conventional farm (vs. not), high within-farm density, and early weaning. The main protective factors for total AMU in sows/sucklings were E. coli vaccination in sows and having boars for estrus detection from own production. Regarding antimicrobial class-specific outcomes, several risk factors overlapped for weaners and sows/sucklings, such as farmer's non-tertiary education, not having free-sow systems during lactation, and conventional farming. An additional risk factor for weaners was having fully slatted floors. For fatteners, the main risk factor for total AMU was PRRS vaccination in sucklings.CONCLUSIONS: Several factors found here to be associated with AMU. Some were known but others were novel, such as farmer's tertiary education, low pig aggression and free-sow systems which were all associated with lower AMU. These factors provide targets for developing tailor-made interventions, as well as an evidence-based selection of features for further causal assessment and mediation analysis.

To improve our understanding of the health impacts of high and low temperatures, epidemiological studies require spatiotemporally resolved ambient temperature (Ta) surfaces. Exposure assessment over various European cities for multi-cohort studies requires high resolution and harmonized exposures over larger spatiotemporal extents. Our aim was to develop daily mean, minimum and maximum ambient temperature surfaces with a 1 × 1 km resolution for Europe for the 2003–2020 period. We used a two-stage random forest modelling approach. Random forest was used to (1) impute missing satellite derived Land Surface Temperature (LST) using vegetation and weather variables and to (2) use the gap-filled LST together with land use and meteorological variables to model spatial and temporal variation in Ta measured at weather stations. To assess performance, we validated these models using random and block validation. In addition to global performance, and to assess model stability, we reported model performance at a higher granularity (local). Globally, our models explained on average more than 81 % and 93 % of the variability in the block validation sets for LST and Ta respectively. Average RMSE was 1.3, 1.9 and 1.7 °C for mean, min and max ambient temperature respectively, indicating a generally good performance. For Ta models, local performance was stable across most of the spatiotemporal extent, but showed lower performance in areas with low observation density. Overall, model stability and performance were lower when using block validation compared to random validation. The presented models will facilitate harmonized high-resolution exposure assignment for multi-cohort studies at a European scale.

This thesis investigates the application of sentiment analysis in predicting stock returns for the companies listed in the OMXS30 index. The recent development of large language models (LLMs), like ChatGPT, has substantially advanced the field of sentiment analysis. This thesis utilizes Meta’s LlaMA 2 LLM for sentiment analysis, while a random forest model is employed to predict monthly stock returns for the subsequent month. These predictions enable the creation of a dynamic long-short portfolio, with its returns evaluated against the OMXS30 index using the capital asset pricing model, the Fama–French three-factor model, and the Carhart four-factor model. The results demonstrate that it is possible to generate excess returns by leveraging LlaMA 2, with sentiment ingrained as a key risk factor. The thesis contributes to the existing literature by examining the Swedish stock market as well as using the open-source LLM LlaMA 2.

The analysis of the global food trade system holds tremendous significance due to substantial increases in food and agricultural production since the latter half of the 20th century. This growth has been fuelled by factors such as population expansion, economic development, technological advancements, and evolving dietary preferences. Simultaneously, we face challenges arising from the imperative to sustainably nourish a growing global population while safeguarding natural resources and ecosystems. Striking the balance between ensuring food supply and mitigating environmental harm has become a critical dilemma. The agri-food system serves as both an influencer of ecological shifts and a susceptible entity to climate destabilization, underscoring the need for sustainable production practices. Within this framework, network analysis emerges as a valuable instrument for comprehending the dynamics of global food trade. It aids in identifying pivotal stakeholders, trade patterns, vital relationships, and pertinent factors. This analysis regards the structure and evolution of the global food trade network. Remarkable insights have been unveiled in this domain. Research has probed into the attributes and development of the trade network concerning a wide spectrum of agricultural products as well as specific commodities. Investigations into the virtual water networks associated with agricultural products have been conducted, alongside the identification of community structures. Community detection proves invaluable in grasping the intricate framework of the global food trade. Indeed, several papers explored trade patterns, the formation of communities, the dynamics of the global food system, and potential related shocks (for instance, see [1], [6],[7],[11]). Unlike the existing literature, this paper presents a multi-step approach that combines a community detection method with a supervised learning methodology. The purpose is to identify communities within the agri-food tr

Machine learning advancements have spurred a revolution in healthcare by making it possible to create prediction models for early disease diagnosis. This report introduces the Multiple Disease Prediction System (MDPS), a state-of-the-art approach that uses machine learning to forecast the likelihood of several diseases based on patient data, including medical history, lifestyle, and demographics. The MDPS addresses the growing difficulties in healthcare by focusing on the early detection of multiple diseases. Some of its crucial components include data preparation, feature selection, model training and disease Detection. Despite advantages like early detection and cost savings, dealing with data privacy, model interpretability, and continuous improvements is essential for MDPS's ethical and efficient usage in healthcare. As a result, MDPS has a significant potential to enhance public health and minimize the difficulties associated with chronic illnesses.

As artificial intelligence becomes more prominent in the state of technology, one substantial aspect of its use is in finance - specifically the stock market. Considering four distinct models: linear regression, neural networks, decision trees, and random forests, this research paper has the primary objective of investigating the efficacies of each model and using that knowledge to accurately predict stock market prices for the next day. This is done by conducting a comparative analysis of the percent errors of each model to determine their accuracy; by establishing a simulation framework leveraging past stock market data and splitting it into subsets for testing and training to recursively evaluate each model, the weights of the models will change based on trends and data to predict the price of the next day. Using this model, a feasible stock market predictor is generated - highlighting the potential of artificial intelligence to guide investment strategies and shape the future of finance. This potential is best displayed through the results of our simulation - making over $15,105 by the end. Accentuating the advantages of various stock market approaches while underscoring the limitations, this research provides valuable insights into the advancement of predictive tools in financial markets - offering a broad overview of how machine learning in stock markets will be implemented and allowing investors to navigate the complexities of AI-stock market dynamics in a high-level, abstracted manner.

UNLABELLED: Small sample sizes and loss of sequencing reads during the microbiome data preprocessing can limit the statistical power of differentiating fresh produce phenotypes and prevent the detection of important bacterial species associated with produce contamination or quality reduction. Here, we explored a machine learning-based k-mer hash analysis strategy to identify DNA signatures predictive of produce safety (PS) and produce quality (PQ) and compared it against the amplicon sequence variant (ASV) strategy that uses a typical denoising step and ASV-based taxonomy strategy. Random forest-based classifiers for PS and PQ using 7-mer hash data sets had significantly higher classification accuracy than those using the ASV data sets. We also demonstrated that the proposed combination of integrating multiple data sets and leveraging a 7-mer hash strategy leads to better classification performance for PS and PQ compared to the ASV method but presents lower PS classification accuracy compared to the feature-selected ASV-based taxonomy strategy. Due to the current limitation of generating taxonomy using the 7-mer hash strategy, the ASV-based taxonomy strategy with remarkably less computing time and memory usage is more efficient for PS and PQ classification and applicable for important taxa identification. Results generated from this study lay the foundation for future studies that wish and need to incorporate and/or compare different microbiome sequencing data sets for the application of machine learning in the area of microbial safety and quality of food. IMPORTANCE: Identification of generalizable indicators for produce safety (PS) and produce quality (PQ) improves the detection of produce contamination and quality decline. However, effective sequencing read loss during microbiome data preprocessing and the limited sample size of individual studies restrain statistical power to identify important features contributing to differentiating PS and PQ phenotypes. We ap

A great amount of financial statements of Swedish companies are registered with the Swedish companies registration office, Bolagsverket, as paper copies. These copies are then scanned and made available to the public as image-PDFs, meaning the financial information contained in them is not digitised and therefore not easily processed in an automated manner. Being able to do so, is valuable however, e.g. to offer automated credit risk assessments or investigate fraud cases through modern technologies. In order to digitise this financial information, the company I work for has developed image recognition algorithms that can create a structured data representation of the financial statements. In some cases however, the image recognition fails at creating an accurate representation of what is written in the financial statement. This poses a data quality issue where further applications onto the digitised financial data might be misconstrued and offer a skewed or outright wrong perspective of the underlying financial situation of a certain company. It is therefore the goal of this thesis to develop a solution that can identify cases in which the image recognition data does not match its true counterpart. The solution developed in this thesis is three-fold. First, the quality of the extracted data is evaluated through the assignment of so-called error labels. Second, a Random Forest classifier is trained to be able to predict these error labels and lastly, a quality score is calculated to offer a suggestion of the best possible representation for each value in a financial statement. It is shown that this approach obtains reasonable results and does indeed beat an existing approach to solving the same problem by a considerable margin. The solution build in this thesis therefore offers a valuable extension to the image recognition program, by allowing for a data quality assessment of the extracted information and therefore increasing the confidence one can have in the digit, In this thesis, I build a solution that can identify whether or not a data point is of good or bad quality. These data points are originally created by a program that extracts information from financial reports. Now, this program works by using image recognition technology and the problem is, that sometimes it simply extracts the wrong information from the financial report. I want to figure out when that happens. A financial report contains numbers that describe a company’s financial situation, usually over the period of a year. For example, it contains the company’s revenue and cost figures. In Sweden, these reports are often registered as paper copies with the financial authorities and then they get scanned into PDFs. Because these PDFs are basically just images, the information contained in them cannot easily be processed with computers. It would be very valuable to be able to do that however, for example to offer automated credit risk assessments on these companies. In order to be able to do that, the company I work for has therefore developed this image recognition program which can extract information from the PDFs and store it in a table format. The central problem of my thesis is that sometimes, the program will not extract the correct information. In the PDF, the revenue value might read 4502 SEK, but the program has extracted the value 8502 SEK instead. I want to be able to identify when that happens to make sure that the extracted information is of high quality. To solve this problem, I build a Machine Learning program that can predict whether an extracted number matches the number written in the PDF or not. As it turns out, this is a decent way to identify the incorrect extractions. This means that my Machine Learning program can be integrated with the image recognition program that the company I work for develops, to improve the extracted information it offers.

This report investigates the multi-label classification problem of expense-related documents, aiming to be classified into nine distinct types of categories. Various machine learning methods including Naive Bayes, Logistic Regression, Support Vector Machines, Linear Discriminant Ana- lysis, k-Nearest Neighbors, Decision Trees, Random Forest, Gradient Boosting, and Recurrent Neural Networks were employed and evaluated. The k-Nearest Neighbors model exhibited the lowest overall performance, likely due to its sensitivity to individual training observations and the impact of outliers in a limited dataset. Intriguingly, k-Nearest Neighbors performed worse with Term Frequency - Inverse Document Frequency (TF-IDF) representation compared to Bag- of-Words (BoW), emphasizing the importance of feature representation in k-Nearest Neighbors models. Among the tested models, multinomial Naive Bayes emerged as one of the top performers, despite being initially considered as a benchmark. Linear models showed similar performance, outperforming k-Nearest Neighbors. Tree-based models revealed that Random Forest and Gradi- ent Boosting outperformed the simpler Decision Trees, attributing their success to the use of multiple learners. Recurrent Neural Networks, especially those with pre-trained weights from Skip-grams, underperformed compared to classical classifiers, most likely due to the small data- set available. Classical models were trained on BoW and TF-IDF representations, using a vocabulary con- sisting of only the top-19 most discriminative features (words) for each category chosen using cross-validation, along with eight featured-engineered variables. Two recurrent neural networks models were employed, one with pre-trained word embeddings from Skip-grams, and one without pre-trained word embeddings. Surprisingly, BoW-based models performed similarly or better than TF-IDF-based models, contrary to the expectation that TF-IDF would give more emphasis on important words. The imb

This thesis, "The Impact of Government Policies on the COVID-19 Pandemic in Scandinavia: A Machine Learning Approach," examines the effects of governmental policies on COVID-19 outcomes in Denmark, Finland, Norway, and Sweden. These countries, while similar in many aspects, adopted varied public health measures in response to the pandemic. Using data pre-processing techniques and machine learning algorithms, including Ordinary Least Squares (OLS) regression, Random Forest, and Support Vector Regression (SVR), this study analyzes the effectiveness of different interventions. The results highlight the critical role of testing, strict health policies and vaccination in managing the pandemic. The findings of this thesis reveals some important roles of testing and governments policies in managing the pandemic. Differences in vaccination and booster programs highlight the varying public health strategies, as booster doses showing more impact in Denmark and Finland than in Norway and Sweden. Sweden had a unique approach, characterized by a positive correlation between case numbers and the stringency of public health measures. Regarding death rates, strict public health measures are associated with lower mortality. Vaccination rates does impact the case numbers, by contributing to less death rates. The distinct patterns observed across the four countries, does reflect the diverse approaches of government policies in combating the effect of the pandemic. Overall, this research highlights the different responses of public health strategies in Scandinavia, and provides some valuable insights for public health policy and response.

On December 13, 2004, the Zero Deforestation Law made deforestation illegal in Eastern Paraguay. Deforestation did not stop, however the law may have decreased deforestation. Most deforestation in this region is conducted to clear land for agriculture. In this dissertation, I conduct three analyses to investigate patterns of deforestation in Eastern Paraguay and the impact of this law on aggregate deforestation, predictors of deforestation, and post-deforestation agricultural land uses.The first analysis investigates drivers of deforestation in Eastern Paraguay before the Zero Deforestation Law. I use satellite-derived deforestation data in a linear probability model to investigate how physical and anthropological land characteristics are correlated with forest loss between 2001 and 2004. I find that physical land characteristics including slope, elevation, soil group, and ecoregion were useful for predicting deforestation. Only some anthropological land characteristics predicted deforestation. Distances to towns or roads were not correlated with deforestation. Proximity to patchy forest cover, such as in areas of existing agricultural clearings, was correlated with more deforestation. The second analysis investigates whether deforestation fell after the Zero Deforestation Law was implemented, and whether the relationship between land characteristics and deforestation changed. I use a linear probability model on a panel dataset of deforestation and land characteristics. My results show that deforestation fell by around a quarter after implementation. In addition, the predictive power of some drivers of deforestation weakened. One relationship reversed. Pre-policy, deforestation was more likely with less tree cover nearby, while post-policy, deforestation was more common near more densely forested areas, possibly indicating a desire to hide. My results cannot be explained by other events that took place during this time, and they provide strong evidence that the Zero

The current lack of continuous and temporal monitoring of Chlorophyll- a in certain regions is evident. There are some platforms that collected this data in the past, without current update. For this reason, a model based on a Random Forest classification was trained using the limnological classification based on chlorophyll- a concentration from SeaWiFS data and extrapolated to Landsat-7. The model achieved an accuracy of 62.29% with a moderate kappa coefficient. This approach allowed the development of a well-performing model capable of efficiently monitoring and updating chlorophyll-a data in a region lacking in-situ sensors for this variable., Peer Reviewed

El nombre de ciberatacs és cada vegada més freqüent, afectant no només a individus sinó també a organitzacions. Això pot portar a nombrosos problemes, impulsant l'estudi i desenvolupament de tècniques per mitigar-los tant com sigui possible. En aquest treball, explorem l'ús de tècniques d'aprenentatge automàtic per crear sistemes intel·ligents capaços d'analitzar amenaces, centrant-nos específicament en amenaces relacionades amb dominis. Es recopilaran i processaran dades per entrenar un model d'aprenentatge automàtic supervisat capaç de classificar si un domini és maliciós o no. Es fa una comparació entre aquestes tècniques i les anteriors que utilitzaven regles predefinides. L'enfoc amb regles predefinides provoca un conflicte entre simplicitat i precisió perquè la sortida és binària. Amb el nou enfoc, volem obtenir una probabilitat de risc entre 0 i 1 en comtes del binàri. La conclusió destaca els resultats superiors obtinguts mitjançant l'ús de tècniques d'aprenentatge automàtic. Aquesta nova tècnica utilitzant el model d'aprenentatge automàtic supervisat mostra un rendiment molt millor que el punt de partida., The number of cyber attacks is increasingly common, affecting not only individuals but also organizations. This can lead to numerous problems, prompting the study and development of techniques to mitigate them as much as possible. In this thesis, we explore the use of machine learning techniques to create intelligent systems capable of analyzing threats, focusing specifically on domain-related threats. Data will be collected and processed to train a supervised machine learning model capable of classifying whether a domain is malicious or not. A comparison is drawn between these techniques and the previous one that used predefined rules. The approach with the predefined rules causes a conflict between simplicity and accuracy because the output is binary. With the new approach we want to output a risk probability between 0 and 1 instead. The conclusion highlights the superior results obtained through the use of machine learning techniques. The approach that used the supervised machine learning model shows way better performance than the baseline.

Climate Integrated Assessment Models (IAMs) typically focus on energy and land use to project emissions, and can be used by simple climate models to project temperature outcomes. However, the decision-makers might also be interested in the reverse perspective: given a desired temperature rise, what is the corresponding energy structure? A climate scenario database includes various models, assumptions, discrete values, and incomplete data. To answer the question, machine learning (ML) techniques, Specifically Random Forest (RF), were employed to create an inverse emulator and predict primary energy sources (fossil, oil, renewable) for the mid-20st century (2050), taking into account temperature projections for the end of the century (2100) with reference to the IPCC AR6 dataset. Two methods were employed to select the emulator's input variables: a systematic and a manual selection approach. The uncertainties in the study, including input, parameter, and implementation uncertainties, were addressed using the Monte Carlo method. Finally, two cases were analyzed in detail to explore the potential of the emulator.

BACKGROUND: Transcatheter aortic valve replacement (TAVR) is an important treatment option for patients with severe symptomatic aortic stenosis. It is important to identify predictors of excellent outcomes (good clinical outcomes, more time spent at home) after TAVR that are potentially amenable to improvement. OBJECTIVES: The purpose of the study was to use machine learning to identify potentially modifiable predictors of clinically relevant patient-centered outcomes after TAVR. METHODS: We used data from 8,332 TAVR cases (January 2016-December 2021) from 21 hospitals to train random forest models with 57 patient characteristics (demographics, comorbidities, surgical risk score, lab values, health status scores) and care process parameters to predict the end point, a composite of parameters that designated an excellent outcome and included no major complications (in-hospital or at 30 days), post-TAVR length of stay of 1 day or less, discharge to home, no readmission, and alive at 30 days. We used recursive feature elimination with cross-validation and Shapley Additive Explanation feature importance to identify parameters with the highest predictive values. RESULTS: The final random forest model retained 29 predictors (15 patient characteristics and 14 care process components); the area under the curve, sensitivity, and specificity were 0.77, 0.67, and 0.73, respectively. Four potentially modifiable predictors with relatively high Shapley Additive Explanation values were identified: type of anesthesia, direct movement to stepdown unit post-TAVR, time between catheterization and TAVR, and preprocedural length of stay. CONCLUSIONS: This study identified four potentially modifiable predictors of excellent outcome after TAVR, suggesting that machine learning combined with hospital-level data can inform modifiable components of care, which could support better delivery of care for patients undergoing TAVR.

Cytochrome P450 (CYP) superfamilien er ansvarlig for ca. 75% af alle metaboliske reaktioner. Langt de fleste lægemidler på markedet metaboliseres af de fem mest fremtrædende CYP isoformer - CYP1A2, CYP2C9, CYP2C19, CYP2D6 og CYP3A4. Hæmning af CYP'er er tæt forbundet med lægemiddelbivirkninger, der kan give alvorlige toksikologiske og farmakologiske konsekvenser. Inden for lægemiddeludvikling er det attraktivt at identificere potentielle CYP-inhibitorer. Ud over eksperimentelle tilgange findes der adskillige beregningsmetoder, især maskinlæringsmodeller, til at hurtigt og pålideligt forudsige CYP inhibition. Formålet med dette projekt er at skabe en maskinlæringsmodel, der kan forudsige, om et stof er en inhibitor for de fem vigtigste CYP’er. I denne rapport var der blevet udviklet modeller med hjælp af Python, der tog brug af Random Forest, Neural Network, Support Vector Machine og en konsensus, der var baseret på flertalsafstemning. Modellerne blev udviklet ved hjælp af et stort datasæt, som indeholdt 85715 forbindelser og blev hentet fra PubChem bioassay databasen. Alle modeller blev valideret med stratified 5-fold cross validation (5-CV), og de endelige modeller blev yderligere evalueret ved hjælp af et eksternt datasæt. Derudover blev modellerne evalueret med oversampling og en lang række forskellige fingerprints, herunder Morgan, MACCS, PubChem og en ”Merge” mellem Morgan og MACCS. Resultaterne viste, at upsampling forbedrede den balanceret forudsigelse for modeller med ubalancerede datasæt. Merge fingerprints viste en bedre præstation for de fleste modeller - dog havde Morgan en sammenlignelig præstation. Fra aftagende rækkefølge var modellerne med den højeste ydeevne SVM/Consensus, RF og NN. Consensus modellen viste ingen signifikant forbedring i forhold til SVM. Derfor blev SVM valgt som de endelige modeller. Area under the receiver operating characteristic curve (AUC) og Ma, The human Cytochrome P450 (CYP) superfamily is responsible for ~75% of all metabolic reactions. The vast majority of drugs on the market are metabolized by the five most prominent CYP isoforms, namely CYP1A2, CYP2C9, CYP2C19, CYP2D6, and CYP3A4. Inhibition of CYPs is closely associated with adverse drug reactions, leading to severe toxicological and pharmacological consequences. In drug discovery, it is desirable to identify potential CYP inhibitors. Besides experimental approaches, there exist numerous computational methods, specifically machine learning models, for fast and reliable prediction of CYP inhibition. The aim of this project is to create a machine learning model able to predict whether a compound is an inhibitor for each of the five major CYPs. In this report, models using Random Forest, Neural Network, Support Vector Machine, and consensus classifier based on majority voting were developed using Python. These models were developed using a large dataset containing 85715 compounds that was extracted from the PubChem bioassay database. All models were validated by stratified 5-fold cross validation (5-CV), and the final models were further evaluated using an external dataset. Moreover, the models were evaluated with oversampling and a wide range of different fingerprints, including Morgan, MACCS, PubChem, and a merge between Morgan and MACCS. Results showed that upsampling improved the balanced prediction for imbalanced datasets. Furthermore, merged fingerprint showed superior performance for most models, although the performance of Morgan is comparable. In descending order of performance, the models with the best performance were SVM/Consensus, RF, and NN. The Consensus model showed no significant improvement over SVM. Therefore, the SVM models were chosen as the final models. Area under the receiver operating characteristic curve (AUC) and Matthews Correlation Coefficien

Cytochrome P450 (CYP) superfamilien er ansvarlig for ca. 75% af alle metaboliske reaktioner. Langt de fleste lægemidler på markedet metaboliseres af de fem mest fremtrædende CYP isoformer - CYP1A2, CYP2C9, CYP2C19, CYP2D6 og CYP3A4. Hæmning af CYP'er er tæt forbundet med lægemiddelbivirkninger, der kan give alvorlige toksikologiske og farmakologiske konsekvenser. Inden for lægemiddeludvikling er det attraktivt at identificere potentielle CYP-inhibitorer. Ud over eksperimentelle tilgange findes der adskillige beregningsmetoder, især maskinlæringsmodeller, til at hurtigt og pålideligt forudsige CYP inhibition. Formålet med dette projekt er at skabe en maskinlæringsmodel, der kan forudsige, om et stof er en inhibitor for de fem vigtigste CYP’er. I denne rapport var der blevet udviklet modeller med hjælp af Python, der tog brug af Random Forest, Neural Network, Support Vector Machine og en konsensus, der var baseret på flertalsafstemning. Modellerne blev udviklet ved hjælp af et stort datasæt, som indeholdt 85715 forbindelser og blev hentet fra PubChem bioassay databasen. Alle modeller blev valideret med stratified 5-fold cross validation (5-CV), og de endelige modeller blev yderligere evalueret ved hjælp af et eksternt datasæt. Derudover blev modellerne evalueret med oversampling og en lang række forskellige fingerprints, herunder Morgan, MACCS, PubChem og en ”Merge” mellem Morgan og MACCS. Resultaterne viste, at upsampling forbedrede den balanceret forudsigelse for modeller med ubalancerede datasæt. Merge fingerprints viste en bedre præstation for de fleste modeller - dog havde Morgan en sammenlignelig præstation. Fra aftagende rækkefølge var modellerne med den højeste ydeevne SVM/Consensus, RF og NN. Consensus modellen viste ingen signifikant forbedring i forhold til SVM. Derfor blev SVM valgt som de endelige modeller. Area under the receiver operating characteristic curve (AUC) og Ma, The human Cytochrome P450 (CYP) superfamily is responsible for ~75% of all metabolic reactions. The vast majority of drugs on the market are metabolized by the five most prominent CYP isoforms, namely CYP1A2, CYP2C9, CYP2C19, CYP2D6, and CYP3A4. Inhibition of CYPs is closely associated with adverse drug reactions, leading to severe toxicological and pharmacological consequences. In drug discovery, it is desirable to identify potential CYP inhibitors. Besides experimental approaches, there exist numerous computational methods, specifically machine learning models, for fast and reliable prediction of CYP inhibition. The aim of this project is to create a machine learning model able to predict whether a compound is an inhibitor for each of the five major CYPs. In this report, models using Random Forest, Neural Network, Support Vector Machine, and consensus classifier based on majority voting were developed using Python. These models were developed using a large dataset containing 85715 compounds that was extracted from the PubChem bioassay database. All models were validated by stratified 5-fold cross validation (5-CV), and the final models were further evaluated using an external dataset. Moreover, the models were evaluated with oversampling and a wide range of different fingerprints, including Morgan, MACCS, PubChem, and a merge between Morgan and MACCS. Results showed that upsampling improved the balanced prediction for imbalanced datasets. Furthermore, merged fingerprint showed superior performance for most models, although the performance of Morgan is comparable. In descending order of performance, the models with the best performance were SVM/Consensus, RF, and NN. The Consensus model showed no significant improvement over SVM. Therefore, the SVM models were chosen as the final models. Area under the receiver operating characteristic curve (AUC) and Matthews Correlation Coefficien

Monitoring and modelling soil organic carbon (SOC) in space and time can help us to better understand soil carbon dynamics and is of key importance to support climate change research and policy. Although machine learning (ML) has attracted a lot of attention in the digital soil mapping (DSM) community for its powerful ability to learn from data and predict soil properties, such as SOC, it is better at capturing soil spatial variation than soil temporal dynamics. By contrast, process-oriented (PO) models benefit from mechanistic knowledge to express physiochemical and biological processes that govern SOC temporal changes. Therefore, integrating PO and ML models seems a promising means to represent physically plausible SOC dynamics while retaining the spatial prediction accuracy of ML models. In this study, a hybrid modelling framework was developed and tested for predicting topsoil SOC stock in space and time for a regional cropland area located in eastern China. In essence, the hybrid model uses predictions of the PO model in unsampled years as additional training data of the ML model, with a weighting parameter assigned to balance the importance of SOC values from the PO model and real measurements. The results indicated that temporal trends of SOC stock modelled by PO and ML models were largely different, while they were notably similar between the PO and hybrid models. Cross-validation showed that the hybrid model had the best performance (RMSE = 0.29 kg m−2), with a 19 % improvement compared with the ML model. We conclude that the proposed hybrid framework not only enhances space-time soil carbon mapping in terms of prediction accuracy and physical plausibility, it also provides insights for soil management and policy decisions in the face of future climate change and intensified human activities.

To enhance our understanding of forest carbon sequestration, climate change mitigation and drought impact on forest ecosystems, the availability of high-resolution annual forest growth maps based on tree-ring width (TRW) would provide a significant advancement to the field. Site-specific characteristics, which can be approximated by high-resolution Earth observation by satellites (EOS), emerge as crucial drivers of forest growth, influencing how climate translates into tree growth. EOS provides information on surface reflectance related to forest characteristics and thus can potentially improve the accuracy of forest growth models based on TRW. Through the modelling of TRW using EOS, climate and topography data, we showed that species-specific models can explain up to 52 % of model variance (Quercus petraea), while combining different species results in relatively poor model performance (R2 = 13 %). The integration of EOS into models based solely on climate and elevation data improved the explained variance by 6 % on average. Leveraging these insights, we successfully generated a map of annual TRW for the year 2021. We employed the area of applicability (AOA) approach to delineate the range in which our models are deemed valid. The calculated AOA for the established forest-type models was 73 % of the study region, indicating robust spatial applicability. Notably, unreliable predictions predominantly occurred in the climate margins of our dataset. In conclusion, our large-scale assessment underscores the efficacy of combining climate, EOS and topographic data to develop robust models for mapping annual TRW. This research not only fills a critical void in the current understanding of forest growth dynamics but also highlights the potential of integrated data sources for comprehensive ecosystem assessments.

Background: Antimicrobial use (AMU) has decreased significantly in Dutch pig farms since 2009. However, this decrease has stagnated recently, with relatively high AMU levels persisting mainly among weaners. The aim of this study was to identify farm-level characteristics associated with: i) total AMU and ii) use of specific antimicrobial classes. Methods: In 2020, cross-sectional data from 154 Dutch pig farms were collected, including information on AMU and farm characteristics. A mixed-effects conditional Random Forest analysis was applied to select the subset of features that was best associated with AMU. Results: The main risk factors for total AMU in weaners were vaccination for PRRS in sucklings, being a conventional farm (vs. not), high within-farm density, and early weaning. The main protective factors for total AMU in sows/sucklings were E. coli vaccination in sows and having boars for estrus detection from own production. Regarding antimicrobial class-specific outcomes, several risk factors overlapped for weaners and sows/sucklings, such as farmer's non-tertiary education, not having free-sow systems during lactation, and conventional farming. An additional risk factor for weaners was having fully slatted floors. For fatteners, the main risk factor for total AMU was PRRS vaccination in sucklings. Conclusions: Several factors found here to be associated with AMU. Some were known but others were novel, such as farmer's tertiary education, low pig aggression and free-sow systems which were all associated with lower AMU. These factors provide targets for developing tailor-made interventions, as well as an evidence-based selection of features for further causal assessment and mediation analysis.

Interspecific interactions, including predator-prey, intraguild predation (IGP), and competition, may drive distribution and habitat use of predator communities. However, elucidating the relative importance of these interactions in shaping predator distributions is challenging, especially in marine communities comprising highly mobile species. We used individual-based models (IBMs) to predict the habitat distributions of apex predators, intraguild (IG) prey, and prey. We then used passive acoustic telemetry to test these predictions in a subtropical marine predator community consisting of eight elasmobranch (i.e., shark and ray) species in Bimini, The Bahamas. IBMs predicted that prey and IG prey will preferentially select habitats based on safety over resources (food), with stronger selection for safe habitat by smaller prey. Elasmobranch space-use patterns matched these predictions. Species with predator-prey and asymmetrical IGP (between apex and small mesopredators) interactions showed the clearest spatial separation, followed by asymmetrical IGP among apex and large mesopredators. Competitors showed greater spatial overlap although with finer-scale differences in microhabitat use. Our study suggests space-use patterns in elasmobranchs are at least partially driven by interspecific interactions, with stronger spatial separation occurring where interactions include predator-prey relationships or IGP., Interspecific interactions, including predator-prey, intraguild predation (IGP), and competition, may drive distribution and habitat use of predator communities. However, elucidating the relative importance of these interactions in shaping predator distributions is challenging, especially in marine communities comprising highly mobile species. We used individual-based models (IBMs) to predict the habitat distributions of apex predators, intraguild (IG) prey, and prey. We then used passive acoustic telemetry to test these predictions in a subtropical marine predator community consisting of eight elasmobranch (i.e., shark and ray) species in Bimini, The Bahamas. IBMs predicted that prey and IG prey will preferentially select habitats based on safety over resources (food), with stronger selection for safe habitat by smaller prey. Elasmobranch space-use patterns matched these predictions. Species with predator-prey and asymmetrical IGP (between apex and small mesopredators) interactions showed the clearest spatial separation, followed by asymmetrical IGP among apex and large mesopredators. Competitors showed greater spatial overlap although with finer-scale differences in microhabitat use. Our study suggests space-use patterns in elasmobranchs are at least partially driven by interspecific interactions, with stronger spatial separation occurring where interactions include predator-prey relationships or IGP.

A lo largo de este trabajo, se explorará la aplicación de diferentes técnicas de aprendizaje automático sobre datos estelares. Se utilizarán como herramientas principales Aladin, para la extracción de los datos, y Python, para su posterior limpieza, manipulación y visualización de resultados haciendo uso de librerías como Numpy, Pandas y Seaborn. Para el desarrollo de los algoritmos, se hará uso de la librería Scikit-Learn. Inicialmente se planteará un modelo para detectar la población candidata a pertenecer a un cúmulo estelar a partir de los datos de Gaia y algoritmos de clustering no supervisados como Kmeans. Posteriormente, se desarrollará exitosamente un modelo de aprendizaje automático para realizar una búsqueda por edad usando la fotometría de Gaia y mediante cortes de paralaje y movimientos propios encontrar poblaciones de estrellas viejas candidatas a pertenecer a los cúmulos estudiados. Se han encontrado 137 candidatos para Messier 67, 402 para NGC 2477 y 261 para NGC 1817, todos ellos cúmulos con una edad superior a los 500 millones de años. Para cúmulos jóvenes como Valparaíso 1 y NGC 3766, el modelo no ha encontrado ningún candidato viejo, como era esperado.

Between the 1940s and the 1980s, 1,2,3-trichloropropane (TCP) was widely applied to agricultural soils as an impurity in soil fumigants to eliminate plant parasitic nematodes. TCP is a suspected human carcinogen that has subsequently leached into groundwater, contaminating numerous wells across North America, Europe, and Asia. In this dissertation, I investigated TCP contamination in groundwater through geospatial mapping and treatment analysis. First, I conducted a systematic review of available treatment technologies in peer-reviewed literature. I found that granular activated carbon (GAC), bioremediation, and chemical reduction using Zero Valent Zinc (ZVZ) are the most effective methods for reducing TCP levels in groundwater. However, the practicality of GAC technology is hindered by its low to moderate adsorption capacity which necessitates frequent media replacement. I also found that the global prevalence of TCP contamination in the hydrosphere needs increased monitoring and modeling. To address the dearth of monitoring data, I developed a predictive model for TCP contamination in California’s Central Valley, one of America’s most important agricultural regions. I used three decision tree machine-learning models to predict TCP contamination in groundwater using various explanatory variables. The Random Forest algorithm emerged as the top performer, with precipitation, redox state, and nitrate levels as the most important predictive variables. To estimate the scale of contamination in California, I performed a mass balance model to estimate that between 110,000 to 4.3 million kg of TCP have accumulated in the California subsurface due to historical fumigant applications. Next, I addressed the scarcity of data on point-of-use treatments and alternative carbon feedstocks for TCP removal. I found through lab experiments that pitcher point-of-use filters can reduce TCP levels by up to 99%. Thus, households can use these low-cost filters to reduce the risk of TCP exposure and improve the safety of their water supply. Regarding alternative carbon feedstocks, I performed a batch isotherm analysis of almond biochar. Almond biochar was less efficient compared to coconut shells and coal-based granular activated carbon sorbents. However, considering the lower energy costs of producing almond biochar and its abundance as a local agricultural byproduct, I recommend a Life Cycle Assessment to evaluate its sustainability. Finally, my dissertation concludes with a science policy analysis that recommends a federal maximum contaminant level for TCP for public health. In summary, my investigations yielded valuable results to guide TCP monitoring and treatment efforts and provide practical information and policy guidance to affected communities and decision-makers.

Machine learning methods and agent-based models enable the optimization of the operation of high capacity facilities. In this paper, we propose a method for automatically extracting and cleaning pedestrian traffic detector data for subsequent calibration of the ingress pedestrian model. The data was obtained from the waiting room traffic of a vaccination center. Walking speed distribution, the number of stops, the distribution of waiting times, and the locations of waiting points were extracted. Of the 9 machine learning algorithms, the random forest model achieved the highest accuracy in classifying valid data and noise. The proposed microscopic calibration allows for more accurate capacity assessment testing, procedural changes testing, and geometric modifications testing in parts of the facility adjacent to the calibrated parts. The results show that the proposed method achieves state-of-the-art performance on a violent-flows dataset. The proposed method has the potential to significantly improve the accuracy and efficiency of input model predictions and optimize the operation of high-capacity facilities.

El resumen en castellano de 150 palabras es el sumario de la información contenida en el trabajo y deberá: (a) indicar los objetivos principales de la investigación; (b) describir las metodologías utilizadas; (c) enumerar los resultados; y (d) enunciar las conclusiones principales. Ha de estar escrito de manera impersonal: ‘el presente trabajo analiza…’.

Assessing groundwater potential for sustainable resource management is critically important. In addressing this concern, this study aims to advance the field by developing an innovative approach for Groundwater potential zone (GWPZ) mapping using advanced techniques, such as FuzzyAHP, FuzzyDEMATEL, and Logistic regression (LR) models. GWPZ was carried out by integrating various primary factors, such as hydrologic, soil permeability, morphometric, terrain distribution, and anthropogenic influences, incorporating twenty-seven individual criteria using multi-criteria decision models along with a hybrid approach for the Subarnarekha River basin, India, in Google earth engine (GEE). The predictive capability of the model was evaluated using a Multi-Collinearity test (VIF <10.0), followed by applying a random forest model, considering the weighted impact of the five primary factors. The hybrid model for GWPZ classification showed that 21.97 % (4256.3 km2) of the area exhibited very high potential, while 11.37 % (2202.1 km2) indicated very low potential for GW in this area. Validation of the groundwater level data from 72 observation wells, performed by the Area under receiver operating characteristic (AUROC) curve technique, yielded values ranging between 75 % and 78 % for different models, underscoring the robust predictability of GWPZ. The hybrid and LR-FuzzyAHP models demonstrated remarkable effectiveness in GWPZ mapping, indicating that the downstream and southern regions boast substantial groundwater potential attributed to alluvial soil and favorable recharge conditions. Conversely, the central part grapples with a scarcity of groundwater. It holds the potential to assist planners and managers in formulating strategies for managing groundwater levels and alleviating the impacts of future droughts.

Climate change is leading to higher temperatures in Sweden, which increases dryness and the risk of wildfires. Although boreal forests are naturally adapted to seasonal fires, more frequent and intense occurrences may lead to tree mortality and changes in biodiversity, and pose risks to humans and wildlife. Existing low spatial resolution products for mapping burned areas may not be suitable for the Swedish boreal forests, which experiences mainly small fires beyond large fires. The study aims to develop an effective approach for mapping burned areas tailed to the Swedish boreal forest context using Sentinel-2 satellite data and advanced machine learning techniques. The random forest algorithm were employed to identify the most important Sentinel-2 derivatives for discriminating burned areas. The top five selected Sentine2 derivatives were then used to train UNet deep learning models for automated burned area mapping. Three image datasets were created based on the top 3, 4, and 5 derivates resulting from the random forest analysis, along with their corresponding burned and unburned reference labels from 2018 to 2023. Subsequently, three UNet models were trained based on these datasets. Comprehensive model evaluation was conducted using precision, recall, F1-Score, and overall accuracy metrics for both burned and unburned area classification. The results demonstrated a consistent improvement in model performance with the inclusion of more input variables. The UNet model trained on the top five Sentinel-2 derivatives achieved the highest accuracy, with a precision of 0,970 and 0,902, an F1-Score of 0,972 and 0,864, and an overall accuracy of 0,940 for burned and unburned area mapping, respectively. This model exhibited robust classification performance across multiple years, accuratelydelineating burned areas of varying sizes and environmental conditions. The proposed approach leverages the strengths of both machine learning and deep learning techniques, enabling accu, Klimatförändringarna leder till förhöjda temperaturer i Sverige, vilket ökar torrhetsgraden och risken för skogsbränder. Trots att boreala skogar är naturligt anpassade till säsongsbundna bränder kan mer frekventa och intensiva bränder leda till träddöd och förändringar i den biologiska mångfalden samt utgöra en risk för människor och vilda djur. Befintliga produkter med låg rumslig upplösning för kartläggning av brända områden kanske inte är lämpliga för de svenska boreala skogarna, som främst upplever små bränder utöver stora bränder. Studien syftar till att utveckla ett effektivt tillvägagångssätt för att kartlägga brända områden anpassade till den svenska boreala skogen med hjälp av Sentinel-2 satellitdata och avancerade maskininlärningstekniker. Random Forest-algoritmen användes för att identifiera de viktigaste Sentinel-2-derivaten för att diskriminera brända områden. De fem bästa utvalda Sentine-2- derivaten användes sedan för att träna UNet djupinlärningsmodeller för automatisk kartläggning av brända områden. Tre bilddataset skapades baserat på de 3, 4 och 5 bästa derivaten från slumpskogsanalysen, tillsammans med deras motsvarande brända och obrända referensetiketter från 2018 till 2023. Därefter utbildades tre UNet-modeller baserat på dessa dataset. Omfattande modellutvärdering genomfördes med hjälp av precision, återkallande, F1-Score och övergripande noggrannhetsmått för både bränd och obränd områdesklassificering. Resultaten visade en konsekvent förbättring av modellens prestanda med införandet av fler ingångsvariabler. UNet-modellen som tränats på de fem bästa Sentinel2-derivaten uppnådde den högsta noggrannheten, med en precision på 0,970 och 0,902, ett F1-poäng på 0,972 och 0,864 och en total noggrannhet på 0,940 för kartläggning av brända respektive obrända områden. Denna modell uppvisade robust klassificeringsprestanda över flera år och avgränsade noggrant brända områden av varierande storlek ochmiljöförhållanden. Det föreslagna tillvägagångssättet

Alzheimer's disease (AD) is a neurodegenerative disorder characterized by cognitive decline, associated with the presence of amyloid-beta plaques and neurofibrillary tangles. Transgenic mouse models, which replicate AD pathologies, are essential for understanding disease mechanisms and testing potential therapies. However, differences in reported behaviours across various studies highlight the need for more advanced methods of analysis. This study aims to develop a machine learning algorithm to accurately classify mice as healthy or transgenic AD mice based on behavioural data from EthoVision video tracking system. Features from EthoVision were statistically analysed to identify those that significantly distinguish between AD and healthy mice. These identified features were subsequently used to train various machine learning models, including Logistic regression, Random forest, Support vector machine and an ensemble method. An analysis of feature importance revealed that elongation, distance travelled, and velocity were among the most critical indicators of AD-related behavioural changes. The algorithm demonstrated optimal performance using Random forest for females with an accuracy of 77% and using an ensemble method for males with an accuracy of 74%. These results highlight the potential of machine learning to improve the accuracy of behavioural analyses in AD research, providing valuable insights for future diagnostics and treatment strategies.

This thesis is a case study conducted at Magna Electronics to explore the use of machinelearning techniques in improving the predictive quality of electronic control unit (ECU)within the automotive manufacturing. This thesis aims to apply interpretable machinelearning methods to predict potential future ECU failures early. With the interpretablemachine learning the goal is to identify predictive variables that lead to ECU failure andwhich can be used as support for decision making.Logistic Regression and Random Forest were chosen as the machine learning methods,which have been used in research of predictive quality and have different levels of interpretability.TreeSHAP was used on the Random Forest as the post-hoc method to furtherunderstand the results. The models’ performances were quantitatively evaluatedthrough metrics such as accuracy and area under precision-recall curve. Subsequently, thebest-performing models were further analyzed using confusion matrices, precision-recallcurves, and horizontal bar charts to assess the impact of predictive variables.The results of this thesis indicated that while Random Forest outperformed Logistic Regression,both models demonstrated limited capability in accurately predicting faulty ECUs,due to the low AUCPR scores. The precision-recall curves suggested performance near randomguess, highlighting the possible variability in parameter impact.This study has also identified significant challenges, such as data imbalance and mislabeling,which may have had a negative effect on the results. Given these issues, the thesisadvises caution in using these results for decision-making. Although, findings of this thesisunderscore the need for a cautious approach to interpreting model outputs, suggestingthat real-world application may require to use different models based on the specific goalsand context of the analysis.

Probability of Default (PD) models can be used for various purposes within the banking and insurance sector. The models classify an observation as the probability of that observation defaulting within a future time span, often a year. Within banks, this type of model is used to calculate risk grades and capital requirements and needs to be accepted for implementation by Finansinspektionen. Due to requirements from Finansinspektionen regarding transparency and explainability, logistic regression is primarily used today when constructing these models. However, there is an interest in exploring how more advanced machine learning models would perform in this field. On behalf of Länsförsäkringar, this study focuses on how three models built using three different machine learning algorithms perform when trained on the same data as Länsförsäkringar's current model. The algorithms used are Random Forest, XGBoost, and Artificial Neural Networks, and the dataset used consists of private customers holding loans between the years 2007 and 2019. In addition, the study also covers current literature in the field, feature analysis, variable selection, and training of hyperparameters for model optimization. The model that performs the best according to the selected performance measures AUC, Brier score and log loss is the XGBoost model, which is in accordance with findings from several previous studies. The transparency and explainability of this model are found to be inferior to that of logistic regression, but the model does not lack transparency altogether. The study suggests further analysis of how these models could be implemented in the field of PD modelling and how the requirements from Finansinspektionen and EU could be interpreted and changed in order to make reality of the implementation machine learning in risk management., Modeller för beräkning av sannolikheten för fallissemang kan användas för olika ändamål inom bank- och försäkringssektorn. De klassificerar en observation som sannolikheten för att den observationen fallerar inom en framtida tidsperiod, ofta ett år. Hos vissa banker används denna typ av modell för att beräkna riskklasser och kapitalkrav och behöver genomgå en process för godkännande hos Finansinspektionen. På grund av krav av Finansinspektionen gällande transparens och förklarbarhet används idag främst logistisk regression vid konstruktion av dessa modeller. Det finns dock ett intresse för att utforska hur mer avancerade maskininlärningsmodeller skulle prestera inom detta område. På uppdrag av Länsförsäkringar fokuserar denna studie på hur tre modeller, byggda på tre olika maskininlärningsalgoritmer, presterar när de tränas på samma data som Länsförsäkringars nuvarande modell. Algoritmerna som används är Random Forest, XGBoost och Artificial Neural Networks, och datsetet som används består av privatkunder med lån mellan åren 2007 och 2019. Dessutom innehåller studien även en litteraturstudie av området, variabelanalys, variabelval och träning av hyperparametrar för att optimera modellprestationen. Den modell som presterar bäst enligt de utvalda prestationsmåtten AUC, Brier score och log loss är XGBoost, vilket stämmer överens med resultat från flera tidigare studier. Transparensen och förklarbarheten hos denna modell har visat sig vara lägre än för logistisk regression, men möjligheten till transparens är inte obefintlig. Studien föreslår ytterligare utredning av hur dessa modeller skulle kunna införlivas inom PD-modellering och hur kraven från Finansinspektionen och EU skulle kunna tolkas och behöva förändras för att användningen av maskininlärning skulle bli verklighet inom riskhantering.

As industries strive toward sustainability, predictive maintenance emerges as a cornerstone technology, especially with increasing data-driven monitoring within Industry 4.0. Kyoto’s thermal energy storage solution, Heatcube, is a promising technology that has the potential to decarbonize industrial heat while contributing to the integration of renewable electricity into the grid. This thesis explored the potential for Kyoto to advance Heatcube’s maintenance strategies by implementing ML models capable of predicting operational states useful for predictive maintenance integration. For this purpose, two types of Feedforward Neural Networks (FNN) were implemented: one for multiclass classification and another for binary classification. Additionally, a binary classification Random Forest (RF) model was implemented. All these models employed the same proxy sensor data from a water pump to compensate for the lack of historical industrial data. The performances of the different model configurations were compared. The conclusion was that the multiclass FNN model showed the worst predictive performance, while the binary RF model demonstrated the best one. However, challenges related to model generalizability remain, highlighting the need for further refinement before the model can be used in industrial applications., När industrier strävar mot hållbarhet framträder förutsägande underhåll blir mer tilltalande, särskilt med ökande datadriven övervakning inom Industry 4.0. Kyotos lösning för lagring av värmeenergi, Heatcube, är en lovande teknik som har potential att koldioxidutlösa industriell värme samtidigt som den bidrar till integreringen av förnybar el i nätet. Denna avhandling undersökte potentialen för Kyoto att främja Heatcubes underhållsstrategier genom att implementera ML modeller som kan förutsäga drifttillstånd användbara för förutsägande underhållsintegration. För detta ändamål implementerades två typer av feedforward neurala nätverk (FNN): en för flerklassklassificering och en annan för binär klassificering. Dessutom implementerades en binär klassificeringsmodell för Random Forest (RF). Alla dessa modeller använde samma sensor data från en vattenpump för att kompensera för bristen på historiska industriella data. Prestanda för de olika modellkonfigurationerna jämfördes. Slutsatsen var att multiklass FNN modellen visade den sämsta prediktiva prestandan, medan den binära RF-modellen visade den bästa. Utmaningar relaterade till modellens generaliserbarhet kvarstår dock, vilket visar på behovet av ytterligare förfining innan modellen kan användas i industriella tillämpningar.

The building sector in Sweden is responsible for 40 % of the final energy use, and space heating largely contributes to the total energy demand. Minimum energy performance standards and building codes are becoming more strict and comprehensive globally. In the European Union there are substantial goals of achieving a higher energy efficiency in the residential building sector. Energy Performance Certificates (EPCs) are a tool that can be used to track and assess building energy in a standardized way, to follow up on the energy efficiency of buildings. Privately owned single-family houses do not need to have a valid EPC, it is only mandatory to perform an assessment when the house is subject to sale, meaning that only about 20 % of single-family houses in Sweden have an EPC. The objective of the study is therefore to develop a data driven model based on existing EPC data, to estimate the annual space heating energy demand of Swedish single family houses that do not have an EPC. Sub-objectives include testing different data-driven machine learning models to identify which algorithm achieves the highest accuracy. Input features such as floor area, construction year, geographical location and heating source were used to make predictions. The results show that out of the investigated models, the Histogram-GBM performed the best when predicting the space heating energy of an individual house, with a RMSE of 24.33 kWh/m2 /year, MAE of 17.50 kWh/m2 /year, MAPE of 24.62 % and R2 -value of 0.736. When evaluating the results of predictions for the entire single-family housing stock, the errors of each individual house are hidden within the entire housing stock, minimizing the error of the total prediction. The best performing model, Histogram GBM, achieved a MBE of 0.59 kWh/prediction, equivalent to an error smaller than 0.005 %., I Sverige står byggnadssektorn för 40 % av den slutliga energianvändningen, och uppvärmning av byggnader är en stor andel av den totala energianvändningen. Minimikrav för energiprestanda och byggregler blir globalt allt mer strikta och omfattande. EU har betydande mål att uppnå högre energieffektivitet för alla typer av bostadshus. Energideklarationer (EPC) är ett verktyg som kan användas för att samla in och bedömma energianvändning i byggnader på ett standardiserat sätt, för att kunna följa upp energieffektivisering av byggnader. Till skillnad från andra byggnadstyper behöver privat ägda småhus inte ha en giltig energideklaration, de är endast obligatoriska att utfärda när huset säljs, vilket betyder att bara 20 % av småhus i Sverige har en energideklaration. Syftet med studien är därför att utveckla en datadriven modell baserat på data från existerande energideklarationer, för att estimera den årliga värmeenergianvändningen i svenska småhus som saknar en energideklaration. Syftet är också att testa olika data-drivna maskininlärningsmodeller för att identifiera vilken som uppnår det lägsta felet. I modellerna har indata såsom golvyta, byggår, geografisk plats och värmekälla använts för att göra estimeringar. Resultaten visar att av de undersökta modellerna var Histogram-GBM den bästa för att estimera värmeanvändning, med ett RMSE på 24.33 kWh/m2 /år, MAE på 17.50 kWh/m2 /år, MAPE på 24.62 % och ett R2 -värde på 0.736. Resultaten visar också att när värmeanvändning estimeras för hela beståndet av småhus, minimeras det totala felet när alla kvarvarande fel summeras. Den bästa modellen, Histogram-GBM uppnådde ett MBE på 0.59 kWh/estimering, vilket motsvarar ett fel mindre än 0.005 %.

Due to global warming and changing weather conditions, landslide hazards are increasingly threatening global populations. This is particularly concerning for Västra Götaland, Sweden, which has experienced significant material losses from several major landslides. Therefore, there is a need for landslide susceptibility maps (LSM) to identify highly susceptible areas for further implementation of detailed investigations. This study compares the accuracy of three GIS-based methods. Spatial Multi Criteria Decision Analysis (SMCDA), Frequency Ratio (FR), and Random Forest (RF) in delineating susceptible areas in a landslide hazard analysis in Västra Götaland. The accuracy of each LSM is evaluated using three validation methods. Area Under the Receiver Operating Characteristics Curve (AUC) using historical landslide events, how each LSM identify the area affected by the Stenungsund landslide as susceptible, and how well each LSM fits the classification of an existing detailed landslide susceptibility map. This study also evaluates eleven factors for their importance in landslide hazard analysis through a comparative analysis between expert opinions and feature importance from the RF method. Results show significant differences between the models in both accuracy and delineation of susceptible areas. The RF model was the most accurate, with an AUC of 81.59, compared with 72.75 for SMCDA and 60.75 for FR. RF identified the Stenungsund area as highly susceptible, while SMCDA and FR indicated lower susceptibility. RF also had the best agreement with the small-scale LSM. Despite FR’s better agreement with the small-scale LSM, SMCDA is considered the second-best method due to its higher AUC and greater similarity to RF compared with FR. Soil type and soil depth were identified as the most important factors, particularly for excluding non-susceptible areas. However, due to the complex geological composition of the study area, soil depth is preferred over soil type. Proximity to

Syftet med detta projekt är att utveckla en prediktiv modell för att förutsäga fartygsvibrationer baserat på historisk mätdata från lastfartyg. Projektet fokuserar på att använda maskininlärningsmetoder för att förutspå amplituden av vibrationer och identifiera de mest relevanta variablerna för modellens prediktiva förmåga. De undersökta metoderna inkluderar random forest, gradient boosting och neurala nätverk. Resultaten visar att random forest-modellen presterar bäst utifrån prestandamåtten: medelkvadratfelet (MSE), genomsnittliga absoluta felet (MAE) och genomsnittliga procentuella felet (MAPE). Projektet bidrar till en djupare förståelse av maskininlärningens användningsområden för en mer hållbar sjöfart. Syftet är att ge ett underlag till att potentiellt reducera vibrationer som påverkar fartygens komfort och strukturella integritet.

The diesel engine in heavy-duty trucks is a complex system with many components working together, and a malfunction in any of these components can impact engine performance and result in increased emissions. Fault detection and diagnosis have therefore become essential in modern vehicles, ensuring optimal performance and compliance with progressively stricter legal requirements. One of the most common faults in a diesel engineis faulty injectors, which can lead to fluctuations in the amount of fuel injected. Detecting these issues is crucial, prompting a growing interest in exploring additional signals beyond the currently used signal to enhance the performance and robustness of diagnosing this fault. In this work, an investigation was conducted to identify signals that correlate with faulty injectors causing over- and underfueling. It was found that the NOx, O2, and exhaust pressure signals are sensitive to this fault and could potentially serve as additional diagnostic signals. With these signals, two different diagnostic methods were evaluated to assess their effectiveness in detecting injector faults. The methods evaluated were data-driven residuals and Random Forest classifier. The data-driven residuals, when combined with the CUSUM algorithm, demonstrated promising results in detecting faulty injectors. The O2 signal proved effective in identifying both fault instances, while NOx and exhaust pressure were more effective at detecting overfueling. The Random Forest classifier also showed good performance in detecting both over- and underfueling. However, it was observed that using a classifier requires more extensive data preprocessing. Two preprocessing methods were employed: integrating previous measurements and calculating statistical measures over a defined time span. Both methods showed promising results, with the latter proving to be the better choice. Additionally, the generalization capabilities of these methods across different operating conditions were

Zur Kartierung landwirtschaftlicher Flächen mithilfe von Satellitenbildern werden häufig Machine Learning Algorithmen angewandt. Ein Großteil der Anwendungen in der Fernerkundung fokussiert sich jedoch auf die Reproduktion von Daten. Gebiete, für die das Training stattgefunden hat, werden somit klassifiziert, ohne die räumlichen und zeitlichen Abhängigkeiten zu bedenken. In dieser Arbeit wurde die Übertragbarkeit des Random Forest Classifiers für die Detektion von Winterweichweizen (Triticum aestivum subsp. aestivum L.) mithilfe von Sentinel-1 und Sentinel-2 Zeitreihen getestet. Bei der Übertragung eines Classifiers werden für die Auswahl der Satellitenbild-aufnahmen meist Kalenderdaten genutzt. Aufbauend auf der Annahme, dass die Witterungsbedingungen einen erheblichen Einfluss auf die Phänologie haben und der Einbezug vergleichbarer Entwicklungsstadien wiederum einen Einfluss auf die Klassifikationsergebnisse hat, wurden drei Verfahren zur Definition der zu vergleichenden Zeitpunkte entwickelt. Die häufig genutzte Entscheidung anhand von Kalenderdaten bezieht die Unterschiede in der Phänologie nicht mit ein. Allerdings können vor allem zwischen verschiedenen Jahren mehrere Wochen zwischen dem Erreichen eines Entwicklungsstadiums liegen. Aus diesem Grund wurden akkumulierte Temperatursummen (Growing Degree Days) aus meteorologischen Daten errechnet und in einem zweiten Verfahren zur Definition der Interpolationszeitpunkte genutzt und die Ergebnisse des Ontogenesemodells SIMONTO in einem dritten Verfahren verwendet. Für drei Gebiete in Niederösterreich und das Jahr 2022 wurden Random Forest Classifier mithilfe dieser Verfahren und mit Sentinel-1 und Sentinel-2 Zeitreihen trainiert und anschließend auf alle drei Gebiete und die Jahre 2019 bis 2022 angewandt. Die Ergebnisse der 180 auf diese Weise durchgeführten Klassifikationen wurden miteinander vergli-chen, um den Einfluss des Einbezugs meteorologischer Daten zu evaluieren. Die Ergebnisse zeigen, dass durch den Ein, To map agricultural areas with the help of satellite images, machine learning algorithms are often used. However, the majority of these applications in remote sensing focus on the reproduction of data. Areas for which training has taken place are thus classified without considering the spatial and temporal dependencies. In this thesis, an investigation on the spatial and temporal transferability of random forest classifiers for the detection of common winter wheat (Triticum aestivum subsp. aestivum L.) using Sentinel-1 and Sentinel-2 time series is conducted. When transferring a classifier, calendar dates are usually used for the selection of satellite images. Based on the assumption that weather conditions have a considerable influence on phenology and that the consideration of comparable development stages in turn has an influence on the classification results, three different methods were used to define the dates to be considered and compared. The frequently used choice of fixed calendar dates does not take differences in phenology into account. However, especially between different years, the interval between reaching different stages of development can be several weeks. For this reason, accumulated temperature sums (growing degree days) were calculated from meteorological data and used in a second method to define the interpolation times and the results of the SIMONTO ontogenesis model were used in the third method. For three areas within Lower Austria and the year 2022, random forest classifiers were trained with Sentinel-1 and Sentinel-2 data interpolated using these different methods and then applied to all three areas for the years 2019 to 2022. The results of the 180 classifications carried out in this way were compared with each other to evaluate the influence of the inclusion of meteorological data. The results indicate that by the inclusion higher overall accuracies and F-scores could be achieved for the classification of common winter wheat in both the, Masterarbeit Universität Innsbruck 2024

Soil organic carbon (SOC) is an important pool in the carbon cycle that is dependent on a wide array of environmental variables. These are known as the soil forming factors: Parent material, climate, biota, topography and time. Which of those are best suited to predict SOC for a given area depends on the environmental conditions in that area (e.g. arid or humid, vegetation type), but is also understood to be dependent on spatial scale. The aim of this thesis was to investigate this scale dependency of the relations between SOC and environmental variables, using the example of Swedish forest soils. I hypothesized that climate variables would be most important at the national scale, texture would not depend on scale and topographic and chemical variables are the most important at smaller scales. Data on soil properties, soil chemistry and vegetation from the Swedish forest soil inventory was combined with topographic and climatic data to train random forest models of soil organic carbon concentrations in different depths and soil organic carbon stocks in the organic layer. These models were created for the whole of Sweden and for parts of Sweden in the same way, enabling the analysis of how the spatial scale influenced variable importances. The nature of the relationships between SOC and environmental variables was investigated with linear regressions. The main finding was that carbon concentrations are in general best predicted by chemical parameters and that the exact parameters depend more on sample depth than scale. Climate variables were more important at national scale as hypothesized, while the hypotheses regarding chemistry, topography and texture could not be confirmed. For carbon stocks, chemical variables were in total slightly less important and lead by Mn. The lower importance of chemical variables in comparison to carbon concentration models is mainly balanced by the higher importance of topographic variables. Overall, scale dependencies of relationships

Flotation separation is the most important upgrading critical raw material technique. Measuring interactions within flotation variables and modeling their metallurgical responses (grade and recovery) is quite challenging on the industrial scale. These challenges are because flotation separation includes several sub-micron processes, and their monitoring won't be possible for the processing plants. Since many flotation plants are still manually operating and maintaining, understanding interactions within operational variables and their effect on the metallurgical responses would be crucial. As a unique approach, this study used the “Conscious Lab” concept for modeling flotation responses of an industrial copper upgrading plant when Potassium Amyl Xanthate substituted the secondary collector (Sodium Ethyl Xanthate) in the process. The main aim is to understand and compare interactions before and after the collector substitution. For the first time, the conscious lab was constructed based on the most advanced explainable artificial intelligence model, Shapley Additive Explanations, and Catboost. Catboost- Shapley Additive Explanations could accurately model flotation responses (less than 2% error between actual and predicted values) and illustrate variations of complex interactions through the substitution. Through a comparative study, Catboost could generate more precise outcomes than other known artificial intelligence models (Random Forest, Support Vector Regression, Extreme Gradient Boosting, and Convolutional Neural Network). In general, substituting Sodium Ethyl Xanthate by Potassium Amyl Xanthate reduced process predictability, although Potassium Amyl Xanthate could slightly increase the copper recovery., Validerad;2024;Nivå 2;2024-06-10 (signyg);Funder: Array (WISE-WASP);Full text license: CC BY