Downscaling satellite soil moisture for landscape applications: A case study in Delaware, USA

Abstract

Study region: Delaware, USA and its surrounding watersheds. Study focus: An ensemble using multiple Kernel K-nearest neighbors (KKNN) models was trained to predict daily grids of SSM at 100-meter resolution based on SSM estimates from the European Space Agency’s Climate Change Initiative Soil Moisture Product, terrain data, soil maps, and local meteorological network data. Estimated SSM was evaluated against independent in situ SSM observations and were investigated for relationships with land cover class and vegetation phenology (i.e., NDVI). New hydrological insights for the region: Downscaled daily mean SSM estimates had lower error in space (27%) and greater predictive performance over time compared to the raw, coarse resolution remotely sensed SSM dataset when calibrated to field observed values. Downscaled SSM identified stronger and more widespread temporal relationships with NDVI than other estimation methods. However, both coarse and fine resolution datasets greatly underestimated SSM in wetland areas. The findings highlight the need for enhanced in situ SSM monitoring across diverse settings to improve landscape-level downscaled SSM. The downscaling methodology developed in this study was able to produce daily SSM estimates, providing a framework that can support future SSM modeling efforts, hydroecological investigations, and agricultural studies in this and other regions around the world when used in conjunction with ground-based monitoring networks.

Keywords

• Soil moisture
• Remote sensing
• Machine learning
• Spatiotemporal
• Soil moisture networks