Soil salinization is an important limiting factor for agricultural and environmental sustainable development. To achieve rapid and accurate identification of soil salt content, a classification model called attention-bidirectional gate recurrent unit recurrent neural network (Att-BiGRU-RNN) is designed, incorporating the fusion of an attention mechanism. In the encoding and decoding modules of the model, BiGRU and RNN structures are used, enabling the extraction of deep spectral features by leveraging the correlation between spectral information in different bands of hyperspectral data. The attention mechanism is introduced to dynamically allocate weight information based on the differences in spectral information, thereby increasing the contribution of important spectral features to the classification model and improving the accuracy of the model. The research area is initially set in Dinge County, Shaanxi Province, China. Field spectroscopy measurements of 120 samples of original and air-dried soils are conducted using a ground-based spectrometer. Different mixed models for estimating soil salt content, including FDT-SVR, FDT-CNN, BiGRU-RNN, and Att-BiGRU-RNN, are compared. The results show that compared to other models, the Att-BiGRU-RNN model optimized by the attention mechanism exhibits the highest prediction accuracy, with a coefficient of determination R2 = 0.932 and a root mean square error RMSE = 0.012. Additionally, the model’s recall curve at different precision levels is obtained to meet the parameter selection requirements under different estimation demands. This method can effectively identify areas with high soil salt content or severe salinization based on portable hyperspectral sensors and unmanned aerial vehicle platforms and statistically analyze the distribution of soil salt content.
Zhao et al. (Thu,) studied this question.