Integrated use of remote sensing techniques and machine learning in assessing soil salinity: current approaches, methods, and future perspectives

Soil salinity, driven by climate change and human activities, threatens global food security across over 412 million hectares, requiring precise monitoring. Traditional methods are accurate but costly and time-consuming. This review synthesizes recent advances integrating remote sensing (RS; UAVs, satellites, hyperspectral) with machine learning (ML; Random Forest, SVM, ANN) for soil salinity assessment. RS supplies large-scale temporal data, while ML captures complex non-linear relationships. Multi-sensor fusion (e.g. optical-radar) and ensemble ML improve accuracy (R2 up to 0.98). Yet limitations remain: reliance on surface data (0–20 cm), spectral interference from vegetation, small training datasets, and poor generalizability. Deep learning (DL) architectures (e.g. U-Net, FCNN) outperform traditional ML for specific tasks, including semantic segmentation of salinity classes (accuracy > 90%) and subsurface EC prediction (R2 up to 0.85). Future work must close the root-zone salinity gap using DL, time-series analysis, and proximal sensing. A Digital Soil Mapping framework incorporating climatic, topographic, and soil covariates is essential for robust, transferable models. Prioritizing large-scale validation and standardized methods will shift salinity assessment from mapping to predictive management, supporting sustainable agriculture and land restoration.