This study aimed to investigate how the combination of texture information and spectral index affects the accuracy of the soil salinity inversion model. Taking the Bianwan Farm in Jiuquan City, Gansu Province, China as the research area, the multi-spectral data and soil salinity data at 0–15 cm, 15–30 cm and 30–50 cm depths in the sampling area under alfalfa coverage were collected, and spectral reflectance and texture features were obtained from a multispectral image. Moreover, the red-edge band was introduced to improve the spectral index, and gray correlation analysis was utilized to screen sensitive features. Five types of alfalfa-covered soil salinity machine learning inversion models based on random forest (RF) and extreme learning machine (ELM) algorithms were constructed, using the salinity index (SIs), vegetation index (VIs), salinity index + vegetation index (SIs + VIs), vegetation index + texture feature (VIs + TFs), and vegetation index + texture index (VIs + TIs). The determination coefficient R2, root-mean-square error (RMSE) and mean absolute error (MAE) were used to evaluate each model’s performance. The results show that the VIs model is more accurate than the SIs and SIs +VIs models. Combining texture information with VIs improves the inversion accuracy, and the VIs + TIs model has the best inversion effect. From the perspective of inversion depth, the inversion effect for 0–15 cm soil salinity was significantly better than that for other depths, and was the best inversion depth under alfalfa cover. The average R2 of the RF model was 10% higher than that of the ELM. The RF algorithm has high inversion accuracy and stability and performs better than ELM. These findings can serve as a theoretical basis for the efficient inversion of soil salinity and management of saline–alkali lands.
Full Text Finder