Transferability of Soil Moisture Model Developed from Sentinel-1 SAR Data

Abstract

Soil moisture serves as one of the important variables in hydrological processes or climate change. The availability of water in agricultural fields is also a critical information for agriculture, where drought events can affect plant production. Tackling with the large areas covering agricultural fields, spatial information, usually derived from remote sensing, needs to be timely developed. Techniques for retrieving soil moisture have been established using multispectral imageries. Nevertheless, weather condition often constraints information retrievals. In that sense, developing the estimation based on Synthetic Aperture Radar (SAR) data, including Sentinel-1, is a necessity for agricultural planning. The main goals of this research were to model soil moisture using only dual-polarized SAR data and to investigate its potential in transfer modeling using subsequent data acquisition with the presence of vegetation. Soil moisture was modeled using random forest in a set of data and was validated using ground data utilizing soil moisture meter. The result showed that random forest yielded overall accuracy about 0.64 with RMSE 1.54. However, the model was inadequate to perform well when exposed to new data, with averaged accuracy about 0.0045 during transfer modelling. This suggested that the model could be overfit into single date data and was insufficient for a more diverse data range. The research concluded that random forest modelling can be used to estimate soil moisture using remote sensing data but may not easily be transferable. The study, however, suggests a further examination into the characteristics of crop canopy above the soil.