Estimating soil salinity in cotton fields using UAV multispectral remote sensing and SSA–SVM optimised machine learning model

This study developed a sparrow search algorithm-optimised support vector machine (SSA–SVM) screening method using unmanned aerial vehicle (UAV)-based multispectral remote sensing imagery to enhance the accuracy and efficiency of soil salinity detection. This approach effectively extracts weak soil salinity signals in mulched farmland. Spectral, salinity, vegetation indices, and texture features are derived from UAV multispectral data. Boruta, recursive feature elimination (RFE), support vector machine (SVM), and SSA–SVM feature selection methods are applied, followed by modelling with a backpropagation neural network (BPNN) and random forest (RF) to generate a salinity map during the reproductive period. SSA–SVM outperforms Boruta, RFE, and SVM in identifying key texture features, significantly improving soil salinity inversion accuracy. Incorporating texture features enhances model predictions of saline soils, increasing accuracy by 5%. The BPNN model outperforms RF, achieving 10.69% higher accuracy and demonstrating superior capability in modelling nonlinear relationships and variable interactions. A 5-cm resolution salinity distribution map effectively captures spatial and temporal variations across fertility stages. This study advances soil salinity monitoring in precision agriculture, with SSA–SVM feature selection offering a methodological breakthrough that enhances model accuracy and provides critical insights for sustainable agriculture and land management in arid regions.