A deep learning-based method for deep soil salinity prediction: considering the driving mechanisms of salinity profiles

Huifang Chen, Jingwei Wu, Chi Xu,
A deep learning-based method for deep soil salinity prediction: considering the driving mechanisms of salinity profiles,
Geoderma,
Volume 464,
2025,
117615,
ISSN 0016-7061,
https://doi.org/10.1016/j.geoderma.2025.117615.
(https://www.sciencedirect.com/science/article/pii/S0016706125004562)
Abstract: Monitoring deep soil salinity is crucial for the effective management of saline soils and the sustainable development of agriculture, particularly in arid and semi-arid regions. Although remote sensing technologies can estimate soil salt content, the measurement depth is generally limited to the top few centimeters of the soil surface (<5 cm). Given that the vertical migration and accumulation of soil salinity are influenced by multiple factors (meteorology, soil texture, etc.), understanding the transfer relationships and driving mechanisms between surface and subsurface salinity is key to predicting deep soil salinity. Therefore, in this study, Hydrus-1D model simulations and scenario analysis were employed to elucidate the transfer dynamics and driving mechanisms between surface and deep soil salinity, and a Fully Connected Neural Network (FCNN) model was developed to predict deep soil salinity. The results showed that deep soil salinity dynamics throughout the soil profile were jointly driven by multiple factors e.g., irrigation, precipitation, and evapotranspiration, with their interactive effects remaining significant even in deeper layers. The lag response of soil salinity to these driving factors varied considerably and exhibited strong depth dependence. The developed model effectively predicted deep soil salinity, with R2 values ranging from 0.44 to 0.79. This study provides important scientific insights for deep soil salinity management and farmland water–salt regulation.
Keywords: Transfer relationships and driving mechanisms; Hydrus-1D; Deep soil salinity; Lag effects; FCNN