Remote Sensing

Indigenous Lands hold back Amazon deforestation, yet Brazil’s buffer-zone protections cover conservation units and not Indigenous territories, exposing their edges to extraction and drug trafficking. Satellite forest-loss records from 2008 to 2024 were compared inside the Vale do Javari Indigenous Land, noted for its tri-border setting and its large population living in voluntary isolation, and in zones out to 200 kilometers. Clearing there surged from 2017 to 2022, in 2022 running nearly 187.5 times the loss inside the territory, and 64% fell within holdings on the national rural land registry, misused to claim public land. It eased in 2023 and 2024 yet stayed well above the previous decade.

Many countries report crop yields only for large administrative units, so models trained on coarse records must be applied at finer scales, where few samples and scale-dependent relationships degrade accuracy. A machine learning model trained on field-level soybean yields in the US Midwest was fine-tuned with Brazilian state-by-year records, a technique called transfer learning, then used to predict municipal yields from 2001 to 2021. Transfer across scales rather than across regions had rarely been tested. Without any municipal training data, explained yield variance rose from 0.29 to 0.44, though transfer added little where fine-scale data were plentiful.

Nitrate pollution in streams is a significant environmental challenge, affecting water quality and ecosystem health. Hydrological connectivity, the way water moves through the landscape, governs nitrate transport and transformation, but measuring it across entire catchments remains a challenge. Spatial patterns of soil moisture can provide key insights into catchment-scale hydrological connectivity, yet they have not been widely integrated into nitrate prediction models. In this study, we developed a deep learning framework trained on daily streamflow data and SMAP-HydroBlocks, a high-resolution satellite-based soil moisture data set. We tested whether deep learning models can predict nitrate concentrations using spatial soil moisture patterns and applied Explainable AI (XAI) methods to identify which areas contribute most to the model predictions. Our findings reveal that soil moisture patterns significantly improve nitrate predictions, with near-stream areas having the highest predictive power. This study highlights the potential of AI-driven approaches to provide spatially detailed insights into nitrate transport, which could support better water quality management and help identify nutrient pollution hotspots at fine spatial scales.

About half the rain falling on the Amazon returns to the air from soil and plants, so its response to clearing and warming bears on rainfall across South America. Existing estimates are too coarse to detect fine-scale signal of deforestation. We used a surface energy balance model on satellite imagery resolved dry-season losses at 30 meters across Brazil’s 1.3 million square kilometer arc of deforestation, from 1990 to 2022. Over remaining forest ET rose 11 ± 1.5% while cleared land lost 41 ± 2.1%; since forest still covers most of the region, the area-wide trend was positive, 10 ± 2.1%. The ET rise is atritubeted to forest responding to warmer, drier air, though continued clearing could reverse it.

Optical satellite methods infer soil wetness from surface reflectance and greenness, filling the gap left by coarse resolution microwave sensors. Calibrated once for a whole region, they are biased toward the dominant land cover; calibrating separately by land cover class offers a way around that. We show that vegetation-type specific calibration on California’s Central Valley cut average soil moisture estimation error from 0.09 to 0.05 cubic meters per cubic meter and resolved individual fields and water pathways.

Nationwide accounts of Brazilian land use have tended to follow a single cover class or biome, leaving unclear how much deforestation fell on long-standing vegetation rather than regrowth. Annual 30-meter satellite land cover maps for 1985-2020 were reprocessed across six Brazil biomes, separating regrown from long-standing vegetation to quantify the country’s land-use transitions at this resolution for the first time. Long-standing vegetation took 72% of clearing and regrowth 17%; of cleared land, 48% was pasture by 2020 and 9% soybean. Pasture peaked nationally in 2006, since giving way to crops, and every biome showed net clearing despite a quarter of cleared land having regrown.

We introduce SMAP-HydroBlocks (SMAP-HB), a high-resolution satellite-based surface soil moisture dataset at an unprecedented 30-m resolution (2015–2019) across the conterminous United States. SMAP-HB was produced by using a scalable cluster-based merging scheme that combines high-resolution land surface modeling, radiative transfer modeling, machine learning, SMAP satellite microwave data, and in-situ observations. We evaluated the resulting dataset over 1,192 observational sites. Its largest benefit of SMAP-HB is the high spatial detail and improved representation of the soil moisture spatial variability and spatial accuracy with respect to SMAP products.

The first hyper-resolution satellite-based surface soil moisture dataset at 30-m resolution over the continental United States

Drought monitoring and yield prediction often rely on coarse-scale hydroclimate data or (infrequent) vegetation indexes that do not always indicate the conditions farmers face in the field. Consequently, decision-making based on these indices can often be disconnected from the farmer reality. Our study focuses on smallholder farming systems in data-sparse developing countries, and it shows how field-scale soil moisture can leverage and improve crop yield prediction and drought impact assessment.