Remote sensing of soil salinity is essential for selecting suitable salt-tolerant crops and improving soil management. Previous research focused mainly on arid regions. Synthetic aperture radar (SAR) data are crucial for wet coasts due to frequent cloudiness, but significant changes in soil moisture and vegetation impede the soil salinity assessment accuracy. This study demonstrated the feasibility of mapping soil salinity on China's wet east coast through combining machine learning and multi-date SAR data. Two field surveys were carried out on June 17 and July 21, 2017. Using recursive feature elimination, this study generated and screened SAR variables derived from Sentinel-1A SAR imagery acquired on 15 individual dates, and developed support vector regression (SVR) based- and random forest regression (RFR) based-soil salinity models, respectively. The SVR models outperformed the RFR models. The SVR models yielded accurate soil salinity estimations for the 2017-06-17 (R2 = 0.98, RPD = 7.01, RMSE = 0.18 dS/m and RRMSE = 6.28%) and 2017-07-21 (R2 = 0.92, RPD = 3.54, RMSE = 0.27 dS/m and RRMSE = 10.17%) field surveys. The two SVR models efficiently mapped the spatial distribution of soil salinity, and clearly exhibited the temporal changes of soil salinity. Therefore, the new entirely image-based framework constructed two accurate soil salinity estimation models. The framework employs the tenfold cross-validation to reduce overfitting and uncertainty, and it has the potential for adoption over other humid saline regions. The operation-friendly framework does not require the information challenging to acquire on site (e.g., soil moisture and surface roughness). In addition, this study highlights the benefit of using multi-date imagery over the single-date image approach (e.g., the image acquired closer to the field survey date).
Keywords: Coast; Estimation; Machine learning; Multi-date Sentinel-1 data; Soil salinity.
© 2025. The Author(s), under exclusive licence to Springer Nature B.V.