Prediction of High-Resolution Soil Moisture Using Multi-source Data and Machine Learning

Soil moisture (SM) stands as a critical meteorological element influencing the dynamic interplay between the land and the atmosphere. Its comprehension, modeling, and examination hold key significance in unraveling this interaction. Information about the surface SM is necessary for predicting crop yield, future disasters, etc. Ground-based SM measurement is accurate but time-consuming and costly. An alternate approach for measuring SM using satellite images is becoming more popular in recent years. Surface SM retrieval with a fine-resolution still poses challenges. The proposed work considers multi-satellite data for predicting high-resolution SM of Oklahoma, USA using multiple Machine Learning (ML) algorithms, such as K-nearest neighbour (KNN), Decision tree (DT), Random forest (RF), and Extra trees regressor (ETR). A high-resolution SM map for the study region is also reported, considering the Soil Moisture Active Passive (SMAP) SM data as the base, Landsat 8 bands, and normalized difference vegetation index (NDVI) data as the reference datasets. The ETR model performed the best with a mean absolute error (MAE) of 0.940 mm, a root mean square error (RMSE) of 1.303mm and a coefficient of determination (R-2) of 0.965. The external validation is carried out with ground-based SM data from the International Soil Moisture Network (ISMN). Both the actual SMAP SM and predicted SM values demonstrate a comparable correlation with the ISMN data.