High-resolution surface soil moisture retrieval: A hybrid machine learning framework integrating change detection and downscaling for precision water management
Soil moisture (SM) is vital for comprehending the hydrological cycle and managing climatic extremes. Fine-scale accurate SM products hold more and more significant value to water management and precise agricultural irrigation. While in-situ measurements provide high accuracy, their limited spatial c...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-08-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225003498 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Soil moisture (SM) is vital for comprehending the hydrological cycle and managing climatic extremes. Fine-scale accurate SM products hold more and more significant value to water management and precise agricultural irrigation. While in-situ measurements provide high accuracy, their limited spatial coverage and high costs necessitate alternative approaches. Remote sensing enables large-scale monitoring; however, satellite-based SM products have relatively lower spatial resolution, making them less suitable for practical applications. This study presents an innovative high-resolution surface soil moisture (SSM) retrieval framework combining machine learning (ML), change detection and downscaling (CD-DS) methods. The procedure is applied over Catalonia, Spain. The framework integrates Sentinel-1 SAR, Sentinel-2 normalized difference vegetation indices (NDVI), and DISPATCH background SSM data to generate 30-m resolution SSM. A novel backscatter difference variable, derived from the change detection method, improves model performance by addressing vegetation. The ML model was trained using in-situ SSM data collected from 2017 to 2021 and validated against independent in-situ measurement datasets. Among the evaluated algorithms, XGBoost model performed best, achieving an R2 of 0.933 and RMSE of 0.023 cm3/cm3. Validation with ground measurements with different landcover types showed an average correlation of 0.63, a ubRMSE of 0.045 cm3/cm3, and minimal bias of 0.024 cm3/cm3. Notably, backscatter difference emerged as the second most important variable in the ML model after background SSM, highlighting its significance in improving SSM retrieval accuracy. Comparisons with data from 54 measurement sites, obtained during a 2015 field campaign, yielded an R value of 0.82, a RMSE of 0.06 cm3/cm3. Temporal analysis revealed strong consistency with in-situ data, capturing seasonal trends and abrupt changes after precipitation and irrigation events. Furthermore, the spatial distribution of SSM is closely aligned with irrigation type maps, showing higher SSM values in irrigated areas and lower values in rainfed regions. This approach delivers precise field-scale SSM estimates, making it a valuable tool for drought monitoring and modern agricultural practices. |
|---|---|
| ISSN: | 1569-8432 |