A deep learning approach for SMAP soil moisture downscaling informed by thermal inertia theory
Deep learning (DL) based methods have recently made remarkable progress in remote sensing (RS) soil moisture (SM) retrieval applications. However, their purely “black box” algorithms suffer from a lack of interpretability, whereas methods based solely on physical mechanisms often underperform in com...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-02-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225000172 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Deep learning (DL) based methods have recently made remarkable progress in remote sensing (RS) soil moisture (SM) retrieval applications. However, their purely “black box” algorithms suffer from a lack of interpretability, whereas methods based solely on physical mechanisms often underperform in complex scenarios. In this study, we attempt to use an SM downscaling approach that integrates thermal inertia (TI) theory with the DenseNet deep network algorithm. This approach provides partial interpretability of the physical mechanisms while utilizing DenseNet’s superior nonlinear learning ability and feature reuse capability for downscaling the Soil Moisture Active Passive (SMAP) satellite product, generating daily 1 km × 1 km SM. A comprehensive assessment of the downscaled results using in-situ SM acquired from 264 International Soil Moisture Network (ISMN) sites densely distributed across the continental U.S. indicated that this downscaling approach had an overall high accuracy, with an average unbiased root mean square error (ubRMSE) of 0.048 m3/m3. In addition, the downscaled SM exhibited marked improvement in spatial details over the original SMAP SM maps, providing clearer land surface features. The proposed SM downscaling approach is a valuable attempt to adopt DL methods with more practical physical meaning and more interpretability in the current RS Big Data era. |
|---|---|
| ISSN: | 1569-8432 |