Estimation of soil salt content (SSC) in the Ebinur Lake Wetland National Nature Reserve (ELWNNR), Northwest China, based on a Bootstrap-BP neural network model and optimal spectral indices

Soil salinity is recognized worldwide as a major threat to agriculture, particularly in arid regions. Producers and decision-makers thus require updated and accurate maps of salinity in agronomical and environmentally relevant regions. The goals of this study were to test various regression models for estimating soil salt content based on hyperspectral data, HJ-CCD images, and Landsat OLI data using; develop optimal band Difference Index (DI), Ratio Index (RI), and Normalization Index (NDI) algorithms for monitoring soil salt content using image and spectral data; and to compare the performances of the proposed models using a Bootstrap-BP neural network model (Bootstrap-BPNN) from different data sources. The results showed that previously published optimal remote sensing parameters can be applied to estimate the soil salt content in the Ebinur Lake Wetland National Nature Reserve (ELWNNR). Optimal band combination indices based on DI, RI, and NDI were developed for different data sources. Then, the Bootstrap-BP neural network model was built using 1000 groups of Bootstrap samples of remote sensing indices (DI, RI and NDI) and soil salt content. When verifying the accuracy of hyperspectral data, the model yields an R2 value of 0.95, a root mean square error (RMSE) of 4.38 g/kg, and a residual predictive deviation (RPD) of 3.36. The optimal model for remote sensing images was the first derivative model of Landsat OLI, which yielded R2 value of 0.91, RMSE of 4.82 g/kg, and RPD of 3.32; these data indicated that this model has a high predictive ability. When comparing the salinization monitoring accuracy of satellite images to that of ground hyperspectral data, the accuracy of the first derivative of the Landsat OLI model was close to that of the hyperspectral parameter model. Soil salt content was inverted using the first derivative of the Landsat OLI model in the study area