Climate Change and Surface Process
Jiali GUO, Yonggang MA, Heng PAN, Na LI, Changning SUN, Qian SUN, Wenchang ZHOU, Yuxuan DANG
Soil salinization is a major factor in arid and semi-arid regions, adversely affecting agricultural production and the ecological environment. Accurately capturing the spatiotemporal distribution of soil salinization has become a key focus in current research across the fields of ecology, geography, and agriculture. In this study, Sentinel-2A imagery from April and July, along with corresponding in-situ measurements of the salinity of surface soil, were utilized to construct soil salinity inversion models for the Ebinur Lake region. Five machine learning algorithms [random forest (RF), support vector regression, decision tree regression, adaptive boosting (Adaboost), and gradient boosting regression tree] and two deep learning methods(deep belief network and fully convolutional network] were employed for this purpose. Variables were selected using the Boruta algorithm to enhance the performance of the model. The results indicate that: (1) In April, the soil salinity exhibited a strong positive correlation with various spectral bands, whereas the overall correlation strength decreased in July. Among multispectral indices, the intensity indices (Int1, Int2), salinity indices (S3, S5, S6, SI, SI1, SI2, SI3), and the ratio index showed strong positive correlations with the soil salinity, whereas the normalized difference index displayed a strong negative correlation. (2) The RF model achieved the highest predictive accuracy in both time periods, with an average R2 and RMSE of 0.72 and 0.13 in April and 0.66, and 0.15 in July, respectively. Therefore, the RF model was identified as the optimal model in this study. Furthermore, in terms of temporal selection, soil salinity inversion in April yielded higher accuracy compared to July, indicating that April is more favorable for soil salinity monitoring in arid regions.