Search

Deep neural networks (DNNs) can face limitations during training for recognition, motivating this study to improve recognition capabilities by optimizing deep learning features for hand gesture image recognition. We propose a novel approach that enhances features from well-trained DNNs using an improved radial basis function (RBF) neural network, targeting recognition within individual gesture categories. We achieve this by clustering images with a self-organizing map (SOM) network to identify optimal centers for RBF training. Our enhanced SOM, employing the Hassanat distance metric, outperforms the traditional K-Means method across a comparative analysis of various distance functions and the expanded number of cluster centers, accurately identifying hand gestures in images. Our training pipeline learns from hand gesture videos and static images, addressing the growing need for machines to interact with gestures. Despite challenges posed by gesture videos, such as sensitivity to hand pose sequences within a single gesture category and overlapping hand poses due to the high similarities and repetitions, our pipeline achieved significant enhancement without requiring time-related training data. We also improve the recognition of static hand pose images within the same category. Our work advances DNNs by integrating deep learning features and incorporating SOM for RBF training.

Background Due to the lack of complexity and variety of stimuli, conventional housing conditions of laboratory mice do not allow these animals to fully express their behavioral repertoire, including manipulative and cognitive activities. Therefore, we designed mechanical puzzles, so-called lockboxes, for mice that can be provided in their home cages. We investigated the impact of the lockbox enrichment on their phenotype and affective state when compared to conventional housing (CH) and super-environmental enrichment (SEE). Methods Young adult female C57BL/6JCrl mice were examined before and after 2-month exposure to the different types of enrichment in a phenotyping test battery, including tests for trait and state anxiety-related behavior, calorimetric measurements, body weight measurements, the analysis of stress hormone metabolite concentrations, and sequential problem-solving abilities with a novel lockbox. At the end of the study, adrenal gland weights were determined and pathohistological evaluation was performed. For all continuous variables, the relative variability was calculated. Results While the different types of enrichment affected trait anxiety-related behavior, neither state anxiety-related behavior nor physiological variables (i.e., bodyweight, resting metabolic rate, stress hormone metabolite concentrations, adrenal gland weights) were influenced. LE improved sequential problem-solving (i.e., solving novel lockboxes) when compared to SEE. Regardless of the housing condition, the relative variability increased in most variables over time, although the coefficient of variation decreased for some variables, especially in animals with access to LE. There was no evidence of toxicopathological effects associated with the material from which the lockboxes were made. Conclusions All lockboxes are available as open-source tool. LE revealed beneficial effects on the affective state of laboratory mice and their performance in solving novel lockboxes. Neither relevant phenotype of the mice nor reproducibility of the data were compromised by LE, similar to SEE. The lockboxes may also be used as novel approach for assessing cognition in mice.

Abstract Simulating the carbon-water fluxes at more widely distributed meteorological stations based on the sparsely and unevenly distributed eddy covariance flux stations is needed to accurately understand the carbon-water cycle of terrestrial ecosystems. We established a new framework consisting of machine learning, determination coefficient (R2), Euclidean distance, and remote sensing (RS), to simulate the daily net ecosystem carbon dioxide exchange (NEE) and water flux (WF) of the Eurasian meteorological stations using a random forest model or/and RS. The daily NEE and WF datasets with RS-based information (NEE-RS and WF-RS) for 3774 and 4427 meteorological stations during 2002–2020 were produced, respectively. And the daily NEE and WF datasets without RS-based information (NEE-WRS and WF-WRS) for 4667 and 6763 meteorological stations during 1983–2018 were generated, respectively. For each meteorological station, the carbon-water fluxes meet accuracy requirements and have quasi-observational properties. These four carbon-water flux datasets have great potential to improve the assessments of the ecosystem carbon-water dynamics.

Abstract The eddy covariance (EC) flux stations have great limitations in the evaluation of the global net ecosystem carbon exchange (NEE) and in the uncertainty reduction due to their sparse and uneven distribution and spatial representation. If the EC stations are linked with widely distributed meteorological stations using machine learning (ML) and remote sensing, it will play a big role in effectively improving the accuracy of the global NEE assessment and reducing uncertainty. In this study, we developed a framework for estimating NEE at meteorological stations. We first optimized the hyperparameters and input variables of the ML model based on the optimization method called an adaptive genetic algorithm. Then, we developed 566 random forest (RF)‐based NEE estimation models by the strategy of spatial leave‐out‐one cross‐validation. We innovatively established the Euclidean distance‐based accuracy projection algorithm of the R square (R2), which could test the accuracy of each model to estimate the NEE of the specific flux at the weather station. Only the model with the highest R2 was selected from the models with a prediction accuracy of R2 > 0.5 for the specific meteorological stations to estimate its NEE. 4674 out of 10,289 weather stations around the world might match at least one of the 566 NEE estimation models with a projected accuracy of R2 > 0.5. The NEE estimation models we screened for the meteorological stations showed a reliable performance and a higher accuracy than the former studies. The NEE values of the most (96.9%) screened meteorological stations around the world are negative (carbon sink) and most (65.3%) of those showed an increasing trend in the mean annual NEE (carbon sink). The NEE dataset produced at the meteorological stations could be used as a supplement to the EC observations and quasi‐observation data to assess the NEE products of the global grid. The NEE dataset is publicly available via the figshare with https://doi.org/10.6084/m9.figshare.20485563.v1.

ABSTRACTThe Mountain-Oasis-Desert System (MODS) is the fundamental landscape component within the vast arid region of Central Asia. Human activities and natural processes cause surface displacement in the MODS, resulting in geoenvironmental issues and hazards. However, surface displacement and its driving mechanisms in the MODS are not understood. In this study, we used the time-series interferometric synthetic aperture radar (InSAR) method to detect deformation in the MODS in north Tianshan Mountain, Xinjiang and investigated its attribution from the geological conditions, groundwater level changes, and climate variability. The results along the vertical gradient indicate that decorrelation is severe in the high mountain areas with ice-marginal environments (over 3600 m above sea level (a.s.l.)) and the forest-covered mid-mountainous belt (1700–2800 m asl), rendering effective detection unfeasible. Dynamic characteristics are identified in subalpine areas with an elevation of 2800–3600 m asl, with maximum horizontal and vertical displacements of −80.2 mm/year and −58.6 mm/year, respectively. The seasonal acceleration is influenced by the combined effects of temperature and precipitation changes. We observed dense subsidence funnels and ground fissures associated with coal mining in the foreland hills (700–1700 m asl). Ground displacements here exceeded −50 mm/year two years after mining activities had ceased. Subsidence has expanded in the oases with elevations of less than 700 m asl, due to extensive groundwater extraction for agricultural irrigation. The correlation coefficient between groundwater level and displacement was 0.89 and 0.45 at wells in the agricultural and desert areas, respectively. The deformation exhibited seasonal variations associated with groundwater level. The deformation remains relatively stable in the surrounding oasis-affected desert areas, with weak fluctuations influenced by seasonal variations in groundwater level.

The dynamics of Gross Primary Productivity (GPP) is key to understand the global carbon cycle. Multiple GPP products are currently available based on remote sensing, Light Use Efficiency model (LUE) or diagnostic biophysical model. However, little knowledge is available on the spatial patterns of the uncertainty of different GPP products and their potential drivers over the Central Asia (CA), a fragile environment for accurate GPP estimation. This study investigates the sensitivity of the 8-day, monthly and yearly GPP uncertainties based on the three-cornered hat (TCH) method and Shapley additive explanation (SHAP) model in terms of vegetation, energy, water, climate and terrain factors in the dryland ecosystem during the 2003–2015 period. Ten GPP products were examined, including one product (FLUXCOM) from machine learning (ML), six products (EC-LUE, FluxSat, LUEopt, MODIS, MuSyQ and VPM) based on the (LUE), two products (GOSIF and NIRv) from satellite-based direct proxies (Proxies) and one product (PML) from the diagnostic biophysical model. The results indicate that the spatial distribution of the ten GPP products in CA showed similar patterns at different time scales, but with values varied at different products and time scales. According to the eddy covariance (EC) observations and the TCH-based uncertainties, the FLUXCOM product showed smaller relative uncertainties than other products. The attribution analysis denotes that the sources of uncertainty of the GPP varied for each product, and thus the improvement strategies for different products should be implemented separately The FLUXCOM should adapt the vegetation- related module to the dryland environment of CA. The LUE model should optimize the LUE parameters for the dryland ecosystem and incorporate the water related variables in the model. The Proxies model should incorporate the water and energy variables (such as soil moisture and radiation) as input data to improve their performance in CA. The diagnostic model should consider the elevation variable as input data, which may improve the performance of the PML in CA. Our results do not only provide an important basis for the selection of GPP products in the study of the carbon cycle in CA, but also offer a new insight into the GPP model development and improvement for the dryland ecosystem.

The dust originating from the extinct lake of the Aral Sea poses a considerable threat to the surrounding communities and ecosystems. The accurate location of these wind erosion areas is an essential prerequisite for controlling sand and dust activity. However, few relevant indicators reported in this current study can accurately describe and measure wind erosion intensity. A novel wind erosion intensity (WEI) of a pixel resolution unit was defined in this paper based on deformation due to the wind erosion in this pixel resolution unit. We also derived the relationship between WEI and soil InSAR temporal decorrelation (ITD). ITD is usually caused by the surface change over time, which is very suitable for describing wind erosion. However, within a pixel resolution unit, the ITD signal usually includes soil and vegetation contributions, and extant studies concerning this issue are considerably limited. Therefore, we proposed an ITD decomposition model (ITDDM) to decompose the ITD signal of a pixel resolution unit. The least-square method (LSM) based on singular value decomposition (SVD) is used to estimate the ITD of soil (SITD) within a pixel resolution unit. We verified the results qualitatively by the landscape photos, which can reflect the actual conditions of the soil. At last, the WEI of the Aral Sea from 23 June 2020, to 5 July 2020 was mapped. The results confirmed that (1) based on the ITDDM model, the SITD can be accurately estimated by the LSM; (2) the Aral Sea is experiencing severe wind erosion; and (3) the middle, northeast, and southeast bare areas of the South Aral Sea are where salt dust storms may occur.

Abstract Blindness is a main threat that affects the daily life activities of any human. Visual prostheses have been introduced to provide artificial vision to the blind with the aim of allowing them to restore confidence and independence. In this article, we propose an approach that involves four image enhancement techniques to facilitate object recognition and localization for visual prostheses users. These techniques are clip art representation of the objects, edge sharpening, corner enhancement and electrode dropout handling. The proposed techniques are tested in a real-time mixed reality simulation environment that mimics vision perceived by visual prostheses users. Twelve experiments were conducted to measure the performance of the participants in object recognition and localization. The experiments involved single objects, multiple objects and navigation. To evaluate the performance of the participants in objects recognition, we measure their recognition time, recognition accuracy and confidence level. For object localization, two metrics were used to measure the performance of the participants which are the grasping attempt time and the grasping accuracy. The results demonstrate that using all enhancement techniques simultaneously gives higher accuracy, higher confidence level and less time for recognizing and grasping objects in comparison to not applying the enhancement techniques or applying pair-wise combinations of them. Visual prostheses could benefit from the proposed approach to provide users with an enhanced perception.

There has been an increasing number of studies on the potential effects of land-use change on the carbon (C) balance. However, few of these studies have focused on arid regions. Cropland in Xinjiang, a typical arid region in China, has expanded dramatically over the last 40 years. This study applied the Carbon Bookkeeping Model to estimate the changes in C stocks resulting from cropland expansion in Xinjiang from 1975 to 2015. The results showed that the area of cropland increased by a factor of ∼1.6. This increase was driven by advancements in agricultural technology and favorable agricultural policies. The increase in cropland area of 2.03 Mha (M = 106) was the result of the clearing of ∼4.09 Mha land for cropland and the conversion of 2.06 Mha cropland to other land cover types. The expansion in cropland resulted in substantial sequestration of C, with that in Xinjiang amounting to 94.24 Tg C (1Tg = 1012g), accounting for 1.4% of the regional C stocks. Land clearing for cropland (LCC) had the greatest contribution to C sequestration in Xinjiang. The rate of increase in C density through LCC was 0.61 Mg C ha−1a−1 and 1.54 Mg C ha−1a−1 from 1975 to 2004 and 2005 to 2015, respectively. C sequestration due to cropland loss (CLO) of 29.40 Tg C was attributed to the expansion of built-up land and afforestation. Sustainable agricultural activities represented by large-scale clearing for cropland were a major C sink in Xinjiang. Therefore, sustainable management of cropland is essential for maintaining a high C density and preventing loss of C to the atmosphere through cropland abandonment in the future.

In this work, we propose a novel denoising technique, the icosahedral mesh denoising network (IMD-Net) for closed genus-0 meshes. IMD-Net is a deep neural network that produces a denoised mesh in a single end-to-end pass, preserving and emphasizing natural object features in the process. A preprocessing step, exploiting the homeomorphism between genus-0 mesh and sphere, remeshes an irregular mesh using the regular mesh structure of a frequency subdivided icosahedron. Enabled by gauge equivariant convolutional layers arranged in a residual U-net, IMD-Net denoises the remeshing invariant to global mesh transformations as well as local feature constellations and orientations, doing so with a computational complexity of traditional conv2D kernel. The network is equipped with carefully crafted loss function that leverages differences between positional, normal and curvature fields of target and noisy mesh in a numerically stable fashion. In a first, two large shape datasets commonly used in related fields, ABC and ShapeNetCore, are introduced to evaluate mesh denoising. IMD-Net’s competitiveness with existing state-of-the-art techniques is established using both metric evaluations and visual inspection of denoised models. Our code is publicly available at https://github.com/jjabo/IMD-Net.

Volcanic sulfur dioxide (SO2) satellite observations are key for monitoring volcanic activity, and for mitigation of the associated risks on both human health and aviation safety. Automatic analysis of this data source, including robust source emission retrieval, is in turn essential for near real-time monitoring applications. We have developed fast and accurate SO2 plume classifier and segmentation algorithms using classic clustering, segmentation and image processing techniques. These algorithms, applied to measurements from the TROPOMI instrument onboard the Sentinel-5 Precursor platform, can help in the accurate source estimation of volcanic SO2 plumes originating from various volcanoes. In this paper, we demonstrate the ability of different pixel classification methodologies to retrieve SO2 source emission with a good accuracy. We compare the algorithms, their strengths and shortcomings, and present plume classification results for various active volcanoes throughout the year 2021, including examples from Etna (Italy), Sangay and Reventador (Ecuador), Sabancaya and Ubinas (Peru), Scheveluch and Klyuchevskoy (Russia), as well as Ibu and Dukono (Indonesia). The developed algorithms, shared as open-source code, contribute to improving analysis and monitoring of volcanic emissions from space.

Convolutional Neural Networks (CNN) are widely employed in the contemporary artificial intelligence systems. However these models have millions of connections between the layers, that are both memory prohibitive and computationally expensive. Employing these models on an embedded mobile application is resource limited with high power consumption and significant bandwidth requirement to access the data from the off-chip DRAM. Reducing the data movement between the on-chip and off-chip DRAM is the main criteria to achieve high throughput and overall better energy efficiency. Our proposed multi-bit accelerator achieves these goals by employing the truncation of the partial sum (Psum) results of the preceding layer before feeding it into the next layer. We exhibit the architecture by inferencing 32-bits for the first convolution layers and sequentially truncate the bits on the MSB/LSB of integer and fractional part without any further training on the original network. At the last fully connected layer, the top-1 accuracy is maintained with the reduced bit width of 14 and top-5 accuracy upto 10-bit width. The computation engine consists of an systolic array of 1024 processing elements (PE). Large CNNs such as AlexNet, MobileNet, SqueezeNet and EfficientNet were used as benchmark CNN model and Virtex Ultrascale FPGA was used to test the architecture. The proposed truncation scheme has 49% power reduction and resource utilization was reduced by 73.25% for LUTs (Look-up tables), 68.76% for FFs (Flip-Flops), 74.60% for BRAMs (Block RAMs) and 79.425% for Digital Signal Processors (DSPs) when compared with the 32 bits architecture. The design achieves a performance of 223.69 GOPS on a Virtex Ultrascale FPGA, the design has a overall gain of 3.63 × throughput when compared to other prior FPGA accelerators. In addition, the overall power consumption is 4.5 × lower when compared to other prior architectures. The ASIC version of the accelerator was designed in 22nm FDSOI CMOS process to achieve a overall throughput of 2.03 TOPS/W with a total power consumption of 791 mW and with a area of 1 mm × 1.2 mm.

3D reconstruction from a single image has many useful applications. However, it is a challenging and ill-posed problem as various candidates can be a solution for the reconstruction. In this paper, we propose a simple, yet powerful, CNN model that generates a point cloud of an object from a single image. 3D data can be represented in different ways. Point clouds have proven to be a common and simple representation. The proposed model was trained end-to-end on synthetic data with 3D supervision. It takes a single image of an object and generates a point cloud with a fixed number of points. An initial point cloud of a sphere shape is used to improve the generated point cloud. The proposed model was tested on synthetic and real data. Qualitative evaluations demonstrate that the proposed model is able to generate point clouds that are very close to the ground-truth. Also, the initial point cloud has improved the final results as it distributes the points on the object surface evenly. Furthermore, the proposed method outperforms the state-of-the-art in solving this problem quantitatively and qualitatively on synthetic and real images. The proposed model illustrates an outstanding generalization to the new and unseen images and scenes.

Quantifying the variation of biophysical parameters and their driving mechanisms is essential for monitoring land surface environmental changes and for understanding the land–atmosphere interaction in the arid region. Due to the complexity of human activities, most researches are limited to climate change, whereas the response analysis of human activities to changes in biophysical parameters are still lacking or not comprehensively considered. Therefore, large biases and uncertainties still exist in the estimates of regional responses to global change. Firstly, we specifically quantified the main human activities related to land use/land cover change (LULCC) in the northern Tianshan Mountains (NTM), and identified the spatiotemporal changes of primary biophysical parameters, including Albedo, leaf area index (LAI), land surface temperature (LST), and Normalized Difference Vegetation Index (NDVI). Then, we tested the performance of the five models used, including multiple linear regression (MLR), random forest (RF), support vector regression (SVR), multi-layer perceptron (MLP), and K-nearest neighbor (KNN). RF outperformed others and was used to quantify and disaggregate the contribution of climate change and human activities to land surface parameters in the NTM. We found a strong spatial heterogeneity in the spatial variation of all biophysical parameters. Except for LST, the annual maximum Albedo, LAI, and NDVI showed a significant increasing trend in the NTM from 2000 to 2019 (p

This paper presents a comprehensive review of the principle and application of deep learning in retinal image analysis. Many eye diseases often lead to blindness in the absence of proper clinical diagnosis and medical treatment. For example, diabetic retinopathy (DR) is one such disease in which the retinal blood vessels of human eyes are damaged. The ophthalmologists diagnose DR based on their professional knowledge, that is labor intensive. With the advances in image processing and artificial intelligence, computer vision-based techniques have been applied rapidly and widely in the field of medical images analysis and are becoming a better way to advance ophthalmology in practice. Such approaches utilize accurate visual analysis to identify the abnormality of blood vessels with improved performance over manual procedures. More recently, machine learning, in particular, deep learning, has been successfully implemented in this area. In this paper, we focus on recent advances in deep learning methods for retinal image analysis. We review the related publications since 1982, which include more than 80 papers for retinal vessels detections in the research scope spanning from segmentation to classification. Although deep learning has been successfully implemented in other areas, we found only 17 papers so far focus on retinal blood vessel segmentation. This paper characterizes each deep learning based segmentation method as described in the literature. Analyzing along with the limitations and advantages of each method. In the end, we offer some recommendations for future improvement for retinal image analysis.

Assessing the well-being of an animal is hindered by the limitations of efficient communication between humans and animals. Instead of direct communication, a variety of parameters are employed to evaluate the well-being of an animal. Especially in the field of biomedical research, scientifically sound tools to assess pain, suffering, and distress for experimental animals are highly demanded due to ethical and legal reasons. For mice, the most commonly used laboratory animals, a valuable tool is the Mouse Grimace Scale (MGS), a coding system for facial expressions of pain in mice. We aim to develop a fully automated system for the surveillance of post-surgical and post-anesthetic effects in mice. Our work introduces a semi-automated pipeline as a first step towards this goal. A new data set of images of black-furred laboratory mice that were moving freely is used and provided. Images were obtained after anesthesia (with isoflurane or ketamine/xylazine combination) and surgery (castration). We deploy two pre-trained state of the art deep convolutional neural network (CNN) architectures (ResNet50 and InceptionV3) and compare to a third CNN architecture without pre-training. Depending on the particular treatment, we achieve an accuracy of up to 99% for the recognition of the absence or presence of post-surgical and/or post-anesthetic effects on the facial expression.

Understanding the impacts of environmental factors on spatial–temporal and large-scale rodent distribution is important for rodent damage prevention. Investigating rat hole density (RHD) is one of the most effective methods to obtain the intensity of rodent damage. However, most of the previous field surveys or UAV-based remote sensing methods can only evaluate small-scale RHD and its influencing factors. However, these studies did not consider large-scale temporal and spatial heterogeneity. Therefore, we collected small-scale and in situ measurement records of RHD on the northern slope of the Tien Shan Mountains in Xinjiang (NTXJ), China, from 1982 to 2015, and then used correlation analysis and Bayesian network (BN) to analyze the environmental impacts on large-scale RHD with satellite remote sensing data such as the GIMMS NDVI product. The results show that the built BN can better quantify causality in the environmental mechanism modeling of RHD. The NDVI and LAI data from satellite remote sensing are important to the spatial–temporal RHD distribution and the mapping in the future. In regions with an elevation higher than 600 m (UPR) and lower than 600 m (LWR) of NTXJ, there are significant differences in the driving mechanism patterns of RHD, which are dependent on the elevation variation. In LWR, vegetation conditions have a weaker impact on RHD than UPR. It is possibly due to the Artemisia eaten by the dominant species Lagurus luteus (LL) in UPR being more sensitive to precipitation and temperature if compared with the Haloxylon ammodendron eaten by the Rhombomys opimus (RO) in LWR. In LWR, grazing intensity is more strongly and positively correlated to RHD than UPR, possibly due to both winter grazing and RO dependency on vegetation distribution; moreover, in UPR, sheep do not feed Artemisia as the main food, and the total vegetation is sufficient for sheep and LL to coexist. Under the different conditions of water availability of LWR and UPR, grazing may affect the ratio of aboveground and underground biomass by photosynthate allocation, thereby affecting the distribution of RHD. In extremely dry years, the RHD of LWR and UPR may have an indirect interactive relation due to changes in grazing systems.

Most of the world’s 1500 active volcanoes are not instrumentally monitored, resulting in deadly eruptions which can occur without observation of precursory activity. The new Sentinel missions are now providing freely available imagery with unprecedented spatial and temporal resolutions, with payloads allowing for a comprehensive monitoring of volcanic hazards. We here present the volcano monitoring platform MOUNTS (Monitoring Unrest from Space), which aims for global monitoring, using multisensor satellite-based imagery (Sentinel-1 Synthetic Aperture Radar SAR, Sentinel-2 Short-Wave InfraRed SWIR, Sentinel-5P TROPOMI), ground-based seismic data (GEOFON and USGS global earthquake catalogues), and artificial intelligence (AI) to assist monitoring tasks. It provides near-real-time access to surface deformation, heat anomalies, SO2 gas emissions, and local seismicity at a number of volcanoes around the globe, providing support to both scientific and operational communities for volcanic risk assessment. Results are visualized on an open-access website where both geocoded images and time series of relevant parameters are provided, allowing for a comprehensive understanding of the temporal evolution of volcanic activity and eruptive products. We further demonstrate that AI can play a key role in such monitoring frameworks. Here we design and train a Convolutional Neural Network (CNN) on synthetically generated interferograms, to operationally detect strong deformation (e.g., related to dyke intrusions), in the real interferograms produced by MOUNTS. The utility of this interdisciplinary approach is illustrated through a number of recent eruptions (Erta Ale 2017, Fuego 2018, Kilauea 2018, Anak Krakatau 2018, Ambrym 2018, and Piton de la Fournaise 2018−2019). We show how exploiting multiple sensors allows for assessment of a variety of volcanic processes in various climatic settings, ranging from subsurface magma intrusion, to surface eruptive deposit emplacement, pre/syn-eruptive morphological changes, and gas propagation into the atmosphere. The data processed by MOUNTS is providing insights into eruptive precursors and eruptive dynamics of these volcanoes, and is sharpening our understanding of how the integration of multiparametric datasets can help better monitor volcanic hazards.

In this paper, we provide the first in-depth evaluation of exploiting Tomographic Synthetic Aperture Radar (TomoSAR) for the task of supervised land-cover classification. Our main contribution is the design of specific TomoSAR features to reach this objective. In particular, we show that classification based on TomoSAR significantly outperforms PolSAR data provided relevant features are extracted from the tomograms. We also provide a comparison of classification results obtained from covariance matrices versus tomogram features as well as obtained by different reference methods, i.e., the traditional Wishart classifier and the more sophisticated Random Forest. Extensive qualitative and quantitative results are shown on a fully polarimetric and multi-baseline dataset from the E-SAR sensor from the German Aerospace Center (DLR).

This paper proposes the use of Stacked Random Forests (SRF) for the classification of Polarimetric Synthetic Aperture Radar images. SRF apply several Random Forest instances in a sequence where each individual uses the class estimate of its predecessor as an additional feature. To this aim, the internal node tests are designed to work not only directly on the complex-valued image data, but also on spatially varying probability distributions and thus allow a seamless integration of RFs within the stacking framework. Experimental results show that the classification performance is consistently improved by the proposed approach, i.e., the achieved accuracy is increased by 4 % and 7 % for one fully- and one dual-polarimetric dataset. This increase only comes at the cost of a linear increased training and prediction time, which is rather limited as the method converges quickly.

A key ingredient to modern data analysis is probability density estimation. However, it is well known that the curse of dimensionality prevents a proper estimation of densities in high dimensions. The problem is typically circumvented by using a fixed set of assumptions about the data, e.g., by assuming partial independence of features, data on a manifold or a customized kernel. These fixed assumptions limit the applicability of a method. In this paper we propose a framework that uses a flexible set of assumptions instead. It allows to tailor a model to various problems by means of 1d-decompositions. The approach achieves a fast runtime and is not limited by the curse of dimensionality as all estimations are performed in 1d-space. The wide range of applications is demonstrated at two very different real world examples. The first is a data mining software that allows the fully automatic discovery of patterns. The software is publicly available for evaluation. As a second example an image segmentation method is realized. It achieves state of the art performance on a benchmark dataset although it uses only a fraction of the training data and very simple features.