Chiang Mai Journal of Science

     Soil salinization is a severe soil degradation process which represents a critical ecological challenge, threatening the sustainable development of agriculture in the Kashgar oasis region of Xinjiang. Therefore, the timely and efficient monitoring and the accurate estimation of soil salinity are highly imperative for the prevention and management of soil salinization. The study presented in this paper involves the development of a new soil salinity inversion model based on the TabNet deep learning algorithm using remote sensing data and environmental variables. The model outperforms common ensemble learning algorithms based on decision trees. This improvement is achieved through the use of attention mechanism and the deep learning architecture in TabNet. In addition, the novelty of the proposed soil salinity inversion model lies in its use of deep learning to construct a inversion model for salinization. The feature variable dataset was initially constructed using the land surface parameters derived from Landsat 8 imagery and other environmental variables influencing soil salinity. This includes data pre-processing for feature selection using the XGBoost model. Separate soil salinity inversion models were developed using XGBoost, LightGBM, CatBoost, CNN and TabNet algorithms, and their performance was compared. The results indicate that TabNet achieved the best predictive performance among the five models, with R² = 0.57, MAE = 8.10, and RMSE = 11.53 on the test dataset. The results of the best performing model, TabNet, and the importance of individual features were subsequently analyzed using SHAP. The effect of some important factors such as groundwater table depth and altitude on salinization is clearly evident. Furthermore, the threshold of groundwater table depth for salinity control in Kashgar was also determined. These results were consistent with the expertise in soil salinization, which further validates the accuracy of the research findings.