Experimental Study on Quantitative Inversion Model of Heavy Metals in Soda Saline-Alkali Soil Based on RBF Neural Network

Soil is an important part of the natural ecosystem and an important material basis for human survival and agricultural production. With the rapid socio-economic development, the high-intensity industrial and agricultural production activities lead to various pollutants such as heavy metals entering the soil through atmospheric deposition and sewage irrigation and continuously enriching in the soil, causing soil salinization and soil heavy metal pollution, both of which are the main causes of global desertification and soil degradation. However, China has very limited arable land, and food security is especially important. Therefore, quickly and accurately invert the heavy metal content of saline land in a large area is an important research topic to ensure food security. This paper establishes a quantitative inversion model of the heavy metal content of manganese (Mn) cobalt (Co) and iron (Fe) in saline land with soil visible-near infrared spectral data in Zhenlai County, Jilin Province. Firstly, Savitzky-Golay smoothing, multiple scattering correction and continuous statistical de-transformation were performed on the raw spectral data respectively; then three spectral indices, namely, ratio (RI), the difference (DI) and normalized (NDI), were constructed based on the pre-processed spectral data, and the model training samples were determined by correlation analysis between the spectral indices and heavy metal contents. The radial basis neural network algorithm was used to model and invert the saline heavy metal contents. Finally, the sensitive band combinations with significant correlation between the spectral indices and the contents of Mn, Co and Fe were determined by the accuracy analysis method of the gradient cycle modeling such as correlation coefficient and the optimal inversion model based on the radial basis neural network algorithm was established for the heavy metal content of saline land. The results show that the correlation coefficients r>0. 70 for Mn, r>0. 80 for Co, and r>0. 80 for Fe. The selected combinations of sensitivity indices are 108, 690, and 31 groups, respectively, and the optimal inversion models R-2 for Mn, Co, and Fe based on the above significant combinations of sensitivity indices are 0. 703 4, 0. 897 6. The RMSEs were 53. 007 3, 1. 059 2 and 0. 363 4, and the average relative accuracies were 88. 64% 90. 36% and 91. 78% respectively. This study provides an effective method for accurate and rapid analysis of heavy metal content in saline soils, which is of great practical importance for achieving soil heavy metal pollution control.