Integrating machine learning with multitemporal remote sensing to quantify spatial soil salinity

Abstract

Soil salinization poses a major threat to global agricultural productivity, degrading over 1.5 billion hectares of farmland worldwide. In Pakistan alone, approximately 5.7 million hectares of arable land nearly 30 % of the country's irrigated area are affected by salinity, leading to substantial crop yield losses. Here, we demonstrate the potential of integrating Remote Sensing (RS) and Machine Learning (ML) to map soil salinity precisely. Using Sentinel-2A and Landsat-8 OLI data, combined with ground measurements of Electrical Conductivity (EC), we trained and validated three ML algorithms: Random Forest (RF), Classification and Regression Tree (CART), and Support Vector Regression (SVR). Through a refined selection process, we identified SI1, SI4, SI5, CRSI, and wetness as the most relevant indicators for soil salinity mapping. Our results show that RF outperforms CART and SVR, achieving R2 values of 0.91 (Sentinel-2A) and 0.86 (Landsat-8). The RF maps accurately depicted saltaffected lands, including the Indus River, swamp areas, agricultural fields, and saltpan areas. We estimate that 179,200 ha (Landsat-8) to 207,300 ha (Sentinel-2A) are affected by salinity. This study highlights the applications and integrations of RS and ML for monitoring soil salinity, providing location-specific real-time information for assessing unproductive land and to develop smart management practices and strategies for effective decision making.