Search

Financiado para publicación en acceso aberto: Universidade da Coruña/CISUG. [Abstract]: During the last years, deep learning techniques have emerged as powerful alternatives to solve biomedical image analysis problems. However, the training of deep neural networks usually needs great amounts of labeled data to be done effectively. This is even more critical in the case of biomedical imaging due to the added difficulty of obtaining data labeled by experienced clinicians. To mitigate the impact of data scarcity, one of the most commonly used strategies is transfer learning. Nevertheless, the success of this approach depends on the effectiveness of the available pre-training techniques for learning from little or no labeled data. In this work, we explore the application of the Context Encoder paradigm for transfer learning in the domain of retinal image analysis. To this aim, we propose several approaches that allow to work with full resolution images and improve the recognition of the retinal structures. In order to validate the proposals, the Context Encoder pre-trained models are fine-tuned to perform two relevant tasks in the domain: vessels segmentation and fovea localization. The experiments performed on different public datasets demonstrate that the proposed Context Encoder approaches allow mitigating the impact of data scarcity, being superior to previous alternatives in this domain. Xunta de Galicia; ED481A 2021/196 Xunta de Galicia; ED481B-2022-025 Xunta de Galicia; ED431C 2020/24 Xunta de Galicia; IN845D 2020/38 Xunta de Galicia; ED431G 2019/01 This research was funded by Instituto de Salud Carlos III, Gov- ernment of Spain, DTS18/00136 research project; Ministerio de Cien- cia e Innovación y Universidades, Government of Spain, RTI2018- 095894-B-I00 research project; Ministerio de Ciencia e Innovación, Government of Spain through the research project with reference PID2019-108435RB-I00; Consellería de Cultura, Educación e Univer- sidade, Xunta de Galicia, Spain through the predoctoral grant contract ref. ED481A 2021/196 and postdoctoral grant contract ref. ED481B- 2022-025; and Grupos de Referencia Competitiva, grant ref. ED431C 2020/24; Axencia Galega de Innovación (GAIN), Spain, Xunta de Galicia, grant ref. IN845D 2020/38; CITIC, Centro de Investigación de Galicia, Spain ref. ED431G 2019/01, receives financial support from Consellería de Educación, Universidade e Formación Profesional, Xunta de Galicia, Spain, through the ERDF (80%) and Secretaría Xeral de Universidades (20%). Funding for open access charge: Universidade da Coruña/CISUG.

International audience; PMIP contributed to the Intergovernmental Panel on Climate Change (IPCC) Assessment Reports (ARs) by placing current climate change into a wider context, evaluating climate model performance in very different climatic states, and constraining climate sensitivity based on paleoclimates.

In ophthalmology, the study of the retinal microcirculation is a key issue in the analysis of many ocular and systemic diseases, like hypertension or diabetes. This motivates the research on improving the retinal vasculature segmentation. Nowadays, Fully Convolutional Neural Networks (FCNs) usually represent the most successful approach to image segmentation. However, the success of these models is conditioned by an adequate selection and adaptation of the architectures and techniques used, as well as the availability of a large amount of annotated data. These two issues become specially relevant when applying FCNs to medical image segmentation as, first, the existent models are usually adjusted from broad domain applications over photographic images, and second, the amount of annotated data is usually scarcer. In this work, we present multimodal transfer learning-based approaches for retinal vascular segmentation, performing a comparative study of recent FCN architectures. In particular, to overcome the annotated data scarcity, we propose the novel application of self-supervised network pretraining that takes advantage of existent unlabelled multimodal data. The results demonstrate that the self-supervised pretrained networks obtain significantly better vascular masks with less training in the target task, independently of the network architecture, and that some FCN architecture advances motivated for broad domain applications do not translate into significant improvements over the vasculature segmentation task.

This is the second of two papers that document the creation of the New European Wind Atlas (NEWA). In Part 1, we described the sensitivity experiments and accompanying evaluation done to arrive at the final mesoscale model setup used to produce the mesoscale wind atlas. In this paper, Part 2, we document how we made the final wind atlas product, covering both the production of the mesoscale climatology generated with the Weather Research and Forecasting (WRF) model and the microscale climatology generated with the Wind Atlas Analysis and Applications Program (WAsP). The paper includes a detailed description of the technical and practical aspects that went into running the mesoscale simulations and the downscaling using WAsP. We show the main results from the final wind atlas and present a comprehensive evaluation of each component of the NEWA model chain using observations from a large set of tall masts located all over Europe. The added value of the WRF and WAsP downscaling of wind climatologies is evaluated relative to the performance of the driving ERA5 reanalysis and shows that the WRF downscaling reduces the mean wind speed bias and spread relative to that of ERA5 from −1.50±1.30 to 0.02±0.78 m s−1. The WAsP downscaling has an added positive impact relative to that of the WRF model in simple terrain. In complex terrain, where the assumptions of the linearized flow model break down, both the mean bias and spread in wind speed are worse than those from the raw mesoscale results.

This is the first of two papers that document the creation of the New European Wind Atlas (NEWA). It describes the sensitivity analysis and evaluation procedures that formed the basis for choosing the final setup of the mesoscale model simulations of the wind atlas. The suitable combination of model setup and parameterizations, bound by practical constraints, was found for simulating the climatology of the wind field at turbine-relevant heights with the Weather Research and Forecasting (WRF) model. Initial WRF model sensitivity experiments compared the wind climate generated by using two commonly used planetary boundary layer schemes and were carried out over several regions in Europe. They confirmed that the most significant differences in annual mean wind speed at 100 m a.g.l. (above ground level) mostly coincide with areas of high surface roughness length and not with the location of the domains or maximum wind speed. Then an ensemble of more than 50 simulations with different setups for a single year was carried out for one domain covering northern Europe for which tall mast observations were available. We varied many different parameters across the simulations, e.g. model version, forcing data, various physical parameterizations, and the size of the model domain. These simulations showed that although virtually every parameter change affects the results in some way, significant changes in the wind climate in the boundary layer are mostly due to using different physical parameterizations, especially the planetary boundary layer scheme, the representation of the land surface, and the prescribed surface roughness length. Also, the setup of the simulations, such as the integration length and the domain size, can considerably influence the results. We assessed the degree of similarity between winds simulated by the WRF ensemble members and the observations using a suite of metrics, including the Earth Mover's Distance (EMD), a statistic that measures the distance between two probability distributions. The EMD was used to diagnose the performance of each ensemble member using the full wind speed and direction distribution, which is essential for wind resource assessment. We identified the most realistic ensemble members to determine the most suitable configuration to be used in the final production run, which is fully described and evaluated in the second part of this study (Dörenkämper et al., 2020).

Presentation at the NEWA Final Workshop in Bilbao. OUTLINE: Uncertainties in model-data comparison Evaluation of WRF regional simulations: Various sources of information: observational uncertainty Realism of WRF in reproducing observed wind variability Sensitivity of WRF wind climatology to different configurations of the model: How can we characterize model spread?

Tis Article contains errors in the Discussion section.

This report explores the quantification of uncertainty in the NEWA production run and probabilistic ensemble. Uncertainty quantification relies on the availability of model simulations and observations. In areas where observations are not available, uncertainty addresses the quantification of the variability resulting from the sensitivity of the model chain tool resulting from the sampling of all the available model configurations tested. Uncertainty can be here regarded as the quantification of model sensitivity or model spread. In areas with availability of observations, the realism of the model ensemble is evaluated by establishing metrics of model-data comparison. Therefore, uncertainty is understood here as the result of the contributions of model sensitivity to different model setups, and of model errors in a model-data comparison framework. The first part of this report explores the uncertainty derived from model sensitivity subjected to the decisions taken regarding the use of different models setups and how these produce variability in model output. The range of this variability has been regarded as spread in model output and has been quantified in various manners. The second part of this report addresses how model performance can be characterised with the data at hand and whether decisions regarding selection of a given model setup for a production run can be taken on the basis of model performance in a variety of situations, using different variables and datasets as observational targets: wind data from tall masts in the Vestas database; wind speed profiles from tall masts and offshore lidars; surface (10-meter) wind data; satellite data and reanalysis outputs.

This report describes the sensitivity studies performed with the mesoscale model WRF in preparation of the mesoscale wind atlas production runs. The objective of this work was to find a modelsetup that is not just a best practice setup but well-founded and based on scientific evaluation.We started with performing some initial sensitivity experiments changing the PBL scheme and the initialisation of the model. The work was distributed among several partners, each conducting the same set of experiments but on a different domain. The objective of this first phase was to ensure that everybody speaks the same language in terms of applying WRF in the context of NEWA. The results were analysed and compared in terms of the mean wind climate. To draw conclusions regarding the quality of the experiments, the results of one domain were compared to tall mast observations. Overall the model showed a good performance with slightly better results for one of the two tested PBL schemes (MYNN) and weekly initialisation of simulations(compared to daily).In the next phase, further sensitivity tests were conducted for one of the previously defined domains, varying a multitude of parameters as e.g. model version, vertical resolution, forcing data and land surface parameterisation. These studies showed that virtually each parameter change is affecting the results in some way, while significant effects on the wind climate are mostly obtained by changes in physical parameterisation e.g. PBL scheme, representation of the land surface and surface roughness. However, also non-physical parameters as the simulation length and the domain size affects the results considerably. The results suggest to use rather small do-mains and not too long simulations (in the order of 1–2 weeks).One of the objectives of NEWA is to create a probabilistic wind atlas, i.e. to provide uncertainty information to the mesoscale wind atlas (see Deliverables D3.1 and D4.4). This will be achieved by generating an ensemble of WRF simulations with different model configurations. While the final ensemble to be run over the complete NEWA domain will only include a few members, a much larger ensemble was run for a smaller sub-domain to find the ensemble members that generate the largest spread and will be used in the final NEWA ensemble. A second objective of this initial large ensemble was to find an optimal setup for the mesoscale production run. Based on the experience gained in the previous sensitivity experiments, a 47-member ensemble was assembled and run. The individual members were compared against each other, as well as against tall mast observations. Different metrics were explored to assess the performance of the members, i.e. not only the usual statistical measures as RMSE, BIAS and correlation but also metrics that compare the wind speed distributions.In the final part of this report we present the ultimate WRF setup for the NEWA production run that was run between August 2018 and March 2019 on the MareNostrum supercomputer in Barcelona.

Spatially resolved climate reconstructions are commonly derived from long instrumental series and proxy data via linear regression based approaches that use the main modes of the climate system. Such reconstructions have been shown to underestimate climate variability and are based upon the assumption that the main modes of climate variability are stationary back in time. Climate models simulate physically consistent climate fields but cannot be taken to represent the real past climate trajectory because of their necessarily simplified scope and chaotic internal variability. Here, we present sensitivity tests of, and a 200-year temperature reconstruction from, the PSR (Proxy Surrogate Reconstruction) method. This method simultaneously capitalizes on the individual strengths of instrumental/proxy data based reconstructions and model simulations by selecting the model states (analogs) that are most similar with proxy/instrumental data available at specific places and specific moments of time. Sensitivity experiments reveal an optimal PSR configuration and indicate that 6,500 simulation years of existing climate models provide a sufficient pool of possible analogs to skillfully reconstruct monthly European temperature fields during the past 200years. Reconstruction verification based upon only seven instrumental stations indicates potential for extensions back in time using sparse proxy data. Additionally the PSR method allows evaluation of single time series, in this case the homogeneity of instrumental series, by identifying inconsistencies with the reconstructed climate field. We present an updated European temperature reconstruction including newly homogenized instrumental records performed with the computationally efficient PSR method that proves to capture the total variance of the target

Ponencia presentada en: XI Congreso de la Asociación Española de Climatología celebrado en Cartagena entre el 17 y el 19 de octubre de 2018. [ES]Este trabajo proporciona un primer análisis de la variabilidad de la temperatura a escalas interanuales y decadales en la Sierra de Guadarrama (SG), un área protegida de alta montaña del Sistema Central en la Península Ibérica. Se utilizan datos observacionales de estaciones situadas en la zona y una simulación regional de alta resolución (1km) del modelo Weather Research and Forecasting (WRF) para analizar la variabilidad de la temperatura durante el periodo 2000-2015. La comparación entre ambos conjuntos de datos permite evaluar el grado de realismo con el que las simulaciones representan la variabilidad observada. Los resultados muestran que el modelo tiene tendencia a infraestimar los valores medios y las anomalías de las temperaturas observacionales en las estaciones de mayor altitud. Se observa un gradiente vertical de temperatura media de -3.95ºC/km, que es sobreestimado por el modelo (-6.52ºC/km).La variabilidad de las anomalías térmicas aumenta con la altitud para las observaciones y, en menor medida, para las simulaciones. Se evalúa el valor añadido que ofrece WRF frente al uso del reanálisis ERA Interim, que proporciona las condiciones iniciales y de contorno a la simulación regional. Se observa un menor sesgo de las temperaturas que proporciona el modelo regional en comparación con el reanálisis. Se lleva a cabo un Análisis de Componentes Principales (PCA) sobre el campo de anomalías de temperatura de WRF para la evaluación de su variabilidad. Este análisis proporciona un primer modo muy dominante que explica el 94% de la varianza total y cuya componente principal muestra una gran correlación con las anomalías observacionales. La variabilidad de la temperatura en la SG muestra una gran relación con la temperatura en el interior de la Península Ibérica y con una gran parte del suroeste de Europa. El patrón de regresión entre la temperatura proporcionada por la simulación y la que procede del reanálisis permite obtener una estimación de la variabilidad de la temperatura en la SG en los últimos 40 años. [EN]This work provides a first assessment of temperature variability at interannual and decadal timescales in the Sierra de Guadarrama (SG), a high mountain protected area of the Central System in the Iberian Peninsula. Observational data from stations located in the area and simulated data from a high-resolution configuration (1 km) of the Weather Research and Forecasting (WRF) model, fed from ERA Interim inputs, are used in order to analyse the temperature variability in the period 2000-2015. The comparison between both datasets allows for the evaluation of the realism of the model simulations. Results show that the model tends to underestimate observational mean temperatures and anomalies at high altitude stations. A linear mean temperature vertical gradient of -3.95ºC/km is observed and overestimated by the model (-6.52ºC/km).The variability of temperature anomalies for both the observations and, to a lesser extent, the simulations increases with height. The added value that WRF offers against the use of ERA Interim is evaluated. Results show that WRF provides a better performance than the reanalysis, as it shows smaller biases with the observational temperature anomalies. A Principal Component Analysis (PCA) is performed over the temperature anomalies field of WRF for the assessment of its variability. This analysis provides a very dominant first mode that explains the 94% of the total variance and whose PC shows a large correlation with the observational anomalies. Temperature variability in the SG shows a large relationship with temperature in the midland Iberian Peninsula and broadly over south-western Europe. The regression patterns between WRF and the reanalysis are calculated in order to obtain an estimate of the temperature variability over the SG during the last 40 years.

The pre-industrial millennium is among the periods selected by the Paleoclimate Model Intercomparison Project (PMIP) for experiments contributing to the sixth phase of the Coupled Model Intercomparison Project (CMIP6) and the fourth phase of the PMIP (PMIP4). The past1000 transient simulations serve to investigate the response to (mainly) natural forcing under background conditions not too different from today, and to discriminate between forced and internally generated variability on interannual to centennial timescales. This paper describes the motivation and the experimental set-ups for the PMIP4-CMIP6 past1000 simulations, and discusses the forcing agents orbital, solar, volcanic, and land use/land cover changes, and variations in greenhouse gas concentrations. The past1000 simulations covering the pre-industrial millennium from 850 Common Era (CE) to 1849 CE have to be complemented by historical simulations (1850 to 2014 CE) following the CMIP6 protocol. The external forcings for the past1000 experiments have been adapted to provide a seamless transition across these time periods. Protocols for the past1000 simulations have been divided into three tiers. A default forcing data set has been defined for the Tier 1 (the CMIP6 past1000) experiment. However, the PMIP community has maintained the flexibility to conduct coordinated sensitivity experiments to explore uncertainty in forcing reconstructions as well as parameter uncertainty in dedicated Tier 2 simulations. Additional experiments (Tier 3) are defined to foster collaborative model experiments focusing on the early instrumental period and to extend the temporal range and the scope of the simulations. This paper outlines current and future research foci and common analyses for collaborative work between the PMIP and the observational communities (reconstructions, instrumental data).