A framework for determining the total salt content of soil profiles using time-series Sentinel-2 images and a random forest-temporal convolution network

Soil salinization causes a deterioration in soil health and threatens crop growth. Rapid identification of salini-zation in farmlands is of great significance to improve soil functions and to maintain sustainable land man-agement. As salt moves in soil profiles during plowing and irrigation, the commonly used protocol for measuring and monitoring salt content in topsoil does not provide a thorough assessment. In order to quantify and comprehensively evaluate the salt content in deep soil, this study developed a novel framework for monitoring total salt content in the soil profile to a depth of 1 m by combining information from time-series satellite images and machine learning. The field experiments were conducted in Alar, Southern Xinjiang, with a total of 120 soil samples and 582 measurements of EM38-MK2 apparent electrical conductivity in 2019 and 2020 to quantify the vertical variation in the salt content. A total of 42 covariates derived from time-series Sentinel-2 images, including 20 salinity indices, 10 soil indices, and 12 vegetation indices were used for modeling salinity in the soil profile. From the total covariates, 22 were selected using the Random Forest. Soil salinity which was modeled using a Temporal Convolution Network in 2019 and 2020 and forecast for 2021. The model effectively revealed the spatial and temporal variability of the salt content in the soil profile with R-2 of 0.71 and 0.65 for 2019 and 2020, respectively. The proposed new framework provides an effective method to estimate the salt content in the soil profile for precision agriculture in arid and semi-arid regions.