Quantitative analysis of SO₄^2- in saline soil under areas disturbed and undisturbed by human using BP Neural Network

SO₄^2- is one of the main ions in saline soil, but the inversion of SO₄^2- ion in soils under different levels of human disturbance has rarely been reported. Moreover, the relationship between soil hyperspectral and soil elements is nonlinear, and the traditional linear partial least squares model (PLSR) has limited inversion accuracy for soil elements. In order to quantitatively analyzed soil SO₄^2- content in saline soil, this study selected the saline soils in areas undisturbed and human-disturbed in Changji Hui Autonomous Prefecture of Xinjiang, to predict SO₄^2- contents based on soil hyperspectral by using BP Neural Network. The original (R) and logarithmic transformed (LogR) hyperspectral were subjected as 0-order, first-order and second-order differential preprocessing, respectively. The hyperspectral reflectivity corresponding to the sensitive band was taken as the input variable of the nonlinear BP neural network model, and the hidden node, learning rate and maximum number of iterations of BP were set as 300, 0.01, and 1 000. The training function was trainscg. The SO₄^2- contents of saline soil in undisturbed area (Area A) and human-disturbed area (Area B) were determined by using the scatter plot of measured and predicted SO₄^2- contents, the fitting effect map, and the BP training process. The prediction accuracy was tested by comparison with PLSR results. The simulation showed that the BP prediction accuracy after the second-order differentiation in the Area A was better than the first-order differential, while it was opposite for the Area B. The inversion accuracy of LogR spectral transformation was better than R for both Area A and Area B. The relative prediction performance (RPD), determination coefficient (R²), root mean square error (RMSE) and iteration number of the optimal BP model were 3.309, 0.906, 0.253 and 8 times in the Area A; and 2.234, 0.844, 0.786 and 45 times in the Area B. It indicated that BP predictive ability was strong for SO₄^2- content in Area A and Area B. However, for the first- and second-order differentials of spectral in the Area A and Area B, the RPD values of the PLSR were 1.4-1.8, and the prediction performance was normal; in the 0-order differential of the Area B, the RPD of PLSR were all less than 1.0, which could not predict SO₄^2- content. Therefore, the BP model can perform effectively quantitative analysis of SO₄^2- in different undisturbed or human-disturbed regions.