Non-Growing Season Surface Soil Salinity Estimation: Integrating Multi-Source Remote Sensing Data and Convolutional Neural Network Models in Arid Agricultural Areas

Soil salinization reduces crop productivity and threatens agricultural sustainability in arid regions. Reliable estimation of farmland soil salinity is therefore essential for salinization monitoring and land management. During the non-growing season, limited crop cover increases soil surface exposure. Existing studies have mainly relied on optical indices, although SAR features and terrain variables have also been used to improve estimation accuracy. However, climatic and soil texture variables have not been fully considered after the harvest of crops in farmland. In addition, traditional machine learning methods have difficulty effectively learning the complex nonlinear relationship between multi-source variables and soil salinity. Therefore, this research proposed a method to estimate soil salinity by integrating multi-source remote sensing data with a deep learning model. This study focused on farmland in the Wei-Ku Oasis of northwestern China during the non-growing season. Six variable combination scenarios were constructed using Sentinel-1/2 data and environmental covariates, including terrain, land surface temperature (LST), and soil texture. Support vector regression (SVR), random forest (RF), and convolutional neural network (CNN) models were developed to estimate soil salinity. The results showed that: (1) integrating optical indices, SAR features, terrain variables, LST, and soil texture achieved the highest estimation accuracy; (2) the CNN showed better overall estimation performance than the traditional machine learning models across 50 random-split experiments (R2 = 0.68 and RMSE = 1.24 dS/m); and (3) optical indices contributed most to the SVR and RF models, whereas environmental variables contributed most to the CNN model in this study. This study proposed a soil salinity estimation framework that integrates multi-source remote sensing data with a deep learning model during the non-growing season. It provides new data support and technical support for soil salinity estimation of farmland in arid regions during the non-growing season.