Prediction of low Zn concentrations in soil from mountainous areas of central Yunnan Province using a combination of continuous wavelet transform and Boruta algorithm

Accurately predicting heavy metal concentration in soils is essential for agricultural production and food safety. The hyperspectral technique provides a feasible method for rapidly determining heavy metal concentration. In this study, we use spectral and Box-Cox transformations to pre-process the spectral and soil Zn concentration data to solve the problem of low prediction accuracy caused by the non-normal distribution of soil Zn concentration and the unobvious spectral characteristics. The characteristic wavelengths were selected using competitive adaptive reweighted sampling (CARS) and Boruta algorithms. Back propagation neural network (BPNN) was used to predict the low Zn concentration in the soil. The results showed that the accuracy of the models was mostly improved after Box-Cox transformation for low Zn concentration; both continuous wavelet transform (CWT) and fractional order differential (FOD) could explore the delicate features of the spectra to different degrees; the Boruta algorithm could select the wavelengths that were more relevant to Zn; the best performance could be obtained by combining CWT and Boruta algorithms, where the L5-Boruta model had the highest prediction accuracy with R-2, RMSE and RPIQ of 0.656, 13.050 mg & BULL;kg(-1) and 2.351, respectively.