A joint approach for estimating vertically continuous soil moisture profiles by combining P-band SAR polarimetry with soil hydrological modeling is proposed. The approach compares the decomposed soil scattering component from remotely sensed P-band SAR observations of NASA’s Airborne Microwave Observatory of Subcanopy and Subsurface (AirMOSS) mission with an ensemble of simulated counterparts based on the hydrological model HYDRUS-1D and the soil scattering model multi-layer small perturbation method (SPM). From the best fit between remote sensing and soil modeling, the most probable soil moisture profile can be retrieved. Estimated soil moisture profiles at individual monitoring stations across the U.S. are compared to in situ measurements, as well as the European ReAnalysis (ERA5) land and AirMOSS L4 products. Pearson’s coefficient of determination between estimated and auxiliary products prove the overall feasibility of the proposed method with respective R2of 0.92, 0.95, and 0.87.

Recent progress in self-supervision has shown that pre-training large neural networks on vast amounts of unsupervised data can lead to substantial increases in generalization to downstream tasks. Such models, recently coined foundation models, have been transformational to the field of natural language processing.Variants have also been proposed for image data, but their applicability to remote sensing tasks is limited.To stimulate the development of foundation models for Earth monitoring, we propose a benchmark comprised of six classification and six segmentation tasks, which were carefully curated and adapted to be both relevant to the field and well-suited for model evaluation. We accompany this benchmark with a robust methodology for evaluating models and reporting aggregated results to enable a reliable assessment of progress. Finally, we report results for 20 baselines to gain information about the performance of existing models.We believe that this benchmark will be a driver of progress across a variety of Earth monitoring tasks.

Advancements in Machine Learning (ML) techniques have revolutionized model development and hypothesis testing in many fields including Earth science. ML frameworks enable fast iterations on numerous model architectures and rely heavily on training data to learn patterns or relationships. Therefore, training data are a key building block of reproducible ML pipelines. These data need to be curated accurately, be representative of their domain, and be accessible openly to advance ML applications in Earth science. In this paper, we discuss the need for open-access and ML-ready training datasets in Earth Sciences.

A method for estimating complex soil permittivity (or moisture) and penetration depth based on SAR decomposition is presented. By combining model- and eigen-based decomposition techniques, SAR observations are separated into single scattering components (from soil & vegetation). The proposed method incorporates a multi-layer rough surface scattering model to simulate the soil scattering contribution. From the decomposed soil scattering component, permittivity and thus penetration depth can be estimated from SAR observations. Results are presented for the AirMOSS campaign within the MOISST site, OK, USA. As first results, a median value of 16.44 + 2.02 for the complex permittivity was estimated from the 19 P-band data takes at the SoilSCAPE in situ station in Canton, OK, USA. Overall, the retrieval results for the real part of the complex permittivity are similar to in situ values at 30 cm soil depth, with differences in respective median values of 3.44. The median of penetration depths at the Canton site is 23.91 cm,

Synthetic Aperture Radar (SAR) data from Sentinel-1 mission is a valuable all-weather observation for various monitoring applications on the land surface. In particular, SAR observations have unique signatures over surface water which makes them appropriate to develop a global surface water monitoring system. Existing global surface water products rely on multispectral observations which have significant shortcomings in cloudy regions. In this study, we present a novel Convolutional Neural Network (CNN) model applied to Sentinel-1 observations at global scale to detect surface water. We used an existing product as training data, and implemented several fine tuning strategies to improve accuracy of the model in places with complex land cover type. We provide an accuracy assessment against the existing product as well as an independent human-labeled dataset. This new surface water product has higher spatial resolution (10 m) compared to the existing products, is not impacted by cloud coverage, and can be run in near-real time to detect any surface water changes.