Abstract

Understanding soil moisture dynamics is crucial for crop growth. The digital mapping of field soil moisture distribution provides valuable information for agricultural water management. The optical satellite data provides fine scale soil moisture information for a region. However, these data are greatly limited due to cloud contamination and revisit period. Despite the reported beneficial effects of spatiotemporal fusion methods, the accurate estimates of high-resolution soil moisture through spatiotemporal fusion data are still unclear, particularly when using Sentinel-2/3 fusion images. This study introduces a new soil moisture estimation framework integrating spatio-temporal spectral information from Sentinel-2/3 fusion images and machine learning algorithm,and thus provide spatiotemporally continuous soil moisture estimation. The framework includes four fusion methods (ESTARRFM, Fit-FC, FSDAF and STFMF) and four machine learning models (PLSR, SVM, RF and GBRT). The feasibility of the framework was validated in the Hetao Irrigation Area of Inner Mongolia, China. The results showed that the Sentinel-2/3 fused image generated by Fit-FC was visually the closest to the true image, followed by ESTARFM, FSDAF, and STFMF. The spatiotemporal fusion-machine learning estimation framework provided reliable estimation for multi-layer (0 ∼ 20, 20 ∼ 40 and 40 ∼ 60 cm) soil water in the irrigation area. The dense time series of soil water generated by the framework facilitated the detection of irrigation events in the irrigated farmland. Our findings highlighted the effectiveness of Sentinel-2/3 fused images in providing high-resolution continuous daily monitoring of farmland soil water on a large scale. These high spatial–temporal resolution time series are valuable for monitoring crop growth and water resource management, contributing to further expanding the application of satellite remote sensing in precision agriculture.