Downscaling soil moisture in regions with high soil heterogeneity: the solution based on ensemble learning with...

Area of Science:

• Earth and Environmental Sciences
• Remote Sensing
• Data Science

Background:

• Soil moisture is crucial for land-atmosphere interactions, agriculture, and water management.
• Existing microwave remote sensing methods struggle with accuracy in areas of high vegetation and soil heterogeneity.
• Accurate, high-resolution soil moisture data is needed for effective resource management.

Purpose of the Study:

• To develop and validate a high-resolution soil moisture retrieval method for regions with complex land cover.
• To compare the effectiveness of Gradient Boosting Decision Tree (GBDT) and Random Forest (RF) models for downscaling soil moisture data.
• To generate 1km-resolution soil moisture products for Jilin Province.

Main Methods:

• Constructed a soil moisture index set using MODIS and elevation data, calculating Pearson correlation coefficient (R) and Maximum Information Coefficient (MIC).
• Developed decision tree models (GBDT and RF) using Bagging and Boosting ensemble learning methods to downscale Soil Moisture Active Passive (SMAP) data.
• Validated the 1km-resolution products using Triple Collocation Analysis (TCA), comparison with coarse/fine resolution maps, and in-situ measurements.

Main Results:

• The GBDT model showed superior performance over RF, with higher R² (0.733 vs 0.649) and lower error variance in TCA validation.
• At the network scale, GBDT achieved R=0.798 and RMSE=0.040, outperforming RF (R=0.662, RMSE=0.044).
• Point-scale validation also favored GBDT with R=0.864 and RMSE=0.029, compared to RF (R=0.833, RMSE=0.039).

Conclusions:

• Both GBDT and RF models reliably downscale soil moisture in Jilin Province.
• The Boosting ensemble learning method, particularly GBDT, demonstrated better soil moisture estimation performance.
• The developed 1km-resolution soil moisture products are valuable for agricultural and water resource applications.