Abstract
Dynamic change of Soil Salinity Content (SSC) is widely used to control soil salinization for high production efficiency in modern agriculture. This study aimed to enhance the correlation between SSC and spectral reflectance under vegetation coverage using Sentinel-2 multispectral remote sensing. Soil samples of salinity and moisture were collected at 100 sampling sites with different depth ranges, including <20, 20-40, and >40-60 cm, in Shahaoqu Irrigation Area, Inner Mongolia, China, from June to August in 2019. The multispectral data of Sentinel-2 satellite was acquired synchronously according to the sample time and location. The specific procedure was as follows. Firstly, a depth decision tree was constructed with the normalized difference vegetation index as a branching criterion, where the best one was then selected to determine the optimal depth range for the SSC retrieval of soil samples. Secondly, a classification decision tree was used to divide the soil samples into different categories, according to the normalized vegetation index and Soil Moisture Content (SMC). As such, the category of each soil sample was determined using the classification decision tree. Thirdly, the optimal spectral combination for each category was calculated to serve as an input variable into the SSC inversion model. Several machine learning models were adopted for the SSC inversion models to monitor the SSC at the optimal depth range from the salinity depth decision tree, including Adaptive boosting algorithm (Adaboost), Partial Least Squares Regression (PLSR), Support Vector Machines (SVM), Gradient Boosting Decision Tree (GBDT), and Random Forest (RF). The results showed that the correlation coefficient between SSC and spectral reflectance was above 0.66, considering the decision tree. In terms of soil depth range, the optimal inversion depth for the SSC under vegetation cover was >40-60 cm, followed by <20 cm, but the SSC inversion model presented some limitations in the middle layer (20-40 cm). Furthermore, the inversion accuracy was ranked in the descending order of RF, GBDT, Adaboost, SVM, and PLSR, where the RF and Adaboost presented similar inversion. Correspondingly, the SSC inversion model using ensemble learning demonstrated a strong generalization ability to achieve the ideal and stable inversion under different application scenarios, compared with the other machine learning. Specifically, the SSC inversion model performed the best using RF, where the coefficient of determination, the root mean square error, the residual predictive interquartile range, and the residual predictive deviations were 0.70, 0.25%, 0.35, and 1.67, respectively. The correlation between SSC and spectral reflectance was 0.38 without considering the decision tree, indicating there was no significance in the SSC inversion model. Considering the decision tree, the coefficient of determination of the SSC inversion model was 0.70, indicating that the decision tree effectively enhanced the sensitivity of spectral reflectance to SSC for the high accuracy, particularly when the vegetation on the surface of the soil. Consequently, this finding can provide a promising potential way to monitor the soil salinization in the irrigated areas during crop growth using multi-spectral satellites in modern agriculture.