Inversion of soil salinity in dry and wet seasons based on multi-source spectral data in Yinbei area of Ningxia, China

Soil salinization is one of the main causes of global desertification and soil degradation. Information data about salinity and alkalinity is essential to the treatment of alkalized soil for preventing its further degradation and sustainable development of agriculture. Soil salinization is often characterized with significant spatiotemporal dynamics. Taking the saline soil in Pingluo County as the research object, which is predominant in the Ningxia Yinbei area of Northwestern China, this study aims to explore the salt content of soil in dry and wet seasons, and then compare the accuracy of local models and global models, further to determine the optimal model for retrieving soil salinity using the hyperspectral and multispectral remote sensing. The specific processing is following, based on hyperspectral and Landsat 8 OLI image data in the dry season (April) and wet season (October), First, the hyperspectral data was resampled to the image band range for matching the two, whereas, the 11 salt indices under the two spectral data were calculated separately. Second, different algorithms including pearson correlation coefficient (PCC), stepwise regression (SR) and gray relational analysis (GRA) were applied for the sensitive band and index screening of the measured and image spectral data in the dry and wet seasons. Finally, the quantitative analysis models for soil salinity were established using the partial least squares regression (PLSR), support vector machine (SVM), ridge regression (RR), BP neural networks (BPNN), and geographically weighted regression (GWR) method, respectively. All these regression models were verified to select the optimal model, after comparing the effects of different input variables and different regression methods on the model precision. The results showed that: 1) The soil of the Yinbei region was strongly salt-affected, and the salt content in the wet and dry season was characterized by the intensity variation, where the variation degree of dry season was much higher than that of the wet season. 2) The resampling bands showed a good correlation with the image bands data under different soil salinity. 3) The SR group model achieved the best inversion effect, whereas, the PC and GC groups indicated advantages and disadvantages in different regression algorithms, after comparing of the R2, RMSE and RPD of the salt salinity inversion model under the three filter variables of PCC, GC and SR. 4) In the five inversion models of soil salinity, the GWR model showed a higher accuracy. The SVM and BPNN algorithm performed similarly in the models, based on different variable groups, whereas, the RR performance was the worst, particularly a serious “overfitting” phenomenon in the PLSR model. The evaluation results demonstrated the superiority of the local regression over the global regression model for soil salinity. The measured GC-GWR model in dry season achieved the best inversion effect, where the values of RP2 and RPD were 0.94 and 4.49, in the wet season, whereas, the imaged PCC-GWR model obtained the best inversion effect, where the values of RP2 and RPD were 0.96 and 4.83. These findings can contribute to tackling the regional land salinization and degradation, as well as identification and prevention of soil salinization in local and similar areas, further to soil salinization monitoring and land reclamation in arid or semi-arid regions outside the current study area.