A Multimodal Deep Learning Approach for Soil Moisture Downscaling Using Remote Sensing and Weather Data

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 3
• Main Authors: Xu, Yijia, Cai, Shuohao, Huang, Jingyi, Liu, Jiangui, Shang, Jiali, Yang, Zhengwei, Zhang, Zhou
• Format: Journal Article
• Language: English
• Published: 01.09.2025
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2025JH000639

Understanding soil moisture (SM) dynamics is crucial for environmental and agricultural applications. While satellite‐based SM products provide extensive coverage, their coarse spatial resolution often fails to capture local SM variability. This study presents a multimodal network (MMNet) that integrates remote sensing and weather data to downscale Soil Moisture Active Passive (SMAP) Level‐4 surface SM. We evaluated the performance of MMNet by comparing it with in situ SM observations from the Soil Climate Analysis Network (SCAN) and the United States Climate Reference Network (USCRN) under three scenarios. The results showed that (a) MMNet trained with on‐site data provided accurate SM estimates over time in withheld years; (b) MMNet demonstrated spatial transferability, capturing SM dynamics in regions with sparse or no in situ measurements; and (c) the integration of snapshot and time‐series data was crucial for maintaining the model's accuracy and generalizability across diverse scenarios. The downscaled SM maps demonstrated its potential for producing high‐resolution temporally and spatially continuous SM estimates, which could further support a broad range of environmental and agricultural applications. Soil moisture (SM), or the water content in soil, is essential for understanding various environmental and agricultural issues, including crop growth, vegetation health, wildfire risks, and flooding potential. While NASA's SMAP satellite provides a broad view of SM, its low resolution (9–36 km) misses important local details, limiting its use for applications requiring finer spatial resolution, such as precision agriculture. To generate higher resolution SM estimates, we introduced a deep learning model called MMNet, which builds relationship between in situ SM data and coarse SMAP SM, along with other SM‐related features like soil properties, land surface images, and recent weather conditions, generating SM estimates at 100 m resolution. We tested MMNet across diverse scenarios and analyzed how different data types, that is, snapshot and time‐series data, contribute to SM estimation. MMNet outperformed existing methods and the original SMAP, effectively capturing SM dynamics and spatial variations in both temporal and spatial extrapolation tests. These results highlight the importance of integrating multiple data types for accurate SM estimation and demonstrate that MMNet offers a valuable alternative to current deep learning models. A multimodal network (MMNet) was developed to integrate remote sensing and weather data for soil moisture downscaling MMNet achieved RMSE in a temporal generalization test, enabling temporal gap filling or extension of in situ measurements It captured soil moisture dynamics in spatial and cross‐region generalization tests, supporting interpolation in data‐scarce areas