Hyperspectral remote sensing can be used to effectively identify contaminated elements in soil. However, in the field of monitoring soil heavy metal pollution, hyperspectral remote sensing has the characteristics of high dimensionality and high redundancy, which seriously affect the accuracy and stability of hyperspectral inversion models. To resolve the problem, a gradient boosting regression tree (GBRT) hyperspectral inversion algorithm for heavy metal (Arsenic (As)) content in soils based on Spearman’s rank correlation analysis (SCA) coupled with competitive adaptive reweighted sampling (CARS) is proposed in this paper. Firstly, the CARS algorithm is used to roughly select the original spectral data. Second derivative (SD), Gaussian filtering (GF), and min-max normalization (MMN) pretreatments are then used to improve the correlation between the spectra and As in the characteristic band enhancement stage. Finally, the low-correlation bands are removed using the SCA method, and a subset with absolute correlation values greater than 0.6 is retained as the optimal band subset after each pretreatment. For the modeling, the five most representative characteristic bands were selected in the Honghu area of China, and the nine most representative characteristic bands were selected in the Daye area of China. In order to verify the generalization ability of the proposed algorithm, 92 soil samples from the Honghu and Daye areas were selected as the research objects. With the use of support vector machine regression (SVMR), linear regression (LR), and random forest (RF) regression methods as comparative methods, all the models obtained a good prediction accuracy. However, among the different combinations, CARS-SCA-GBRT obtained the highest precision, which indicates that the proposed algorithm can select fewer characteristic bands to achieve a better inversion effect, and can thus provide accurate data support for the treatment and recovery of heavy metal pollution in soils.

Soil total arsenic (TAs) contamination caused by human activities—such as mining, smelting, and agriculture—is a problem of global concern. Visible/near-infrared (VNIR), X-ray fluorescence spectroscopy (XRF), and laser-induced breakdown spectroscopy (LIBS) do not need too much sample preparation and utilization of chemicals to evaluate total arsenic (TAs) concentration in soil. VNIR with hyperspectral imaging has the potential to predict TAs concentration in soil. In this study, 59 soil samples were collected from the Daye City mining area of China, and hyperspectral imaging of the soil samples was undertaken using a visible/near-infrared hyperspectral imaging system (wavelength range 470–900 nm). Spectral preprocessing included standard normal variate (SNV) transformation, multivariate scatter correction (MSC), first derivative (FD) preprocessing, and second derivative (SD) preprocessing. Characteristic bands were then identified based on Spearman’s rank correlation coefficients. Four regression models were used for the modeling prediction: partial least squares regression (PLSR) (R² = 0.71, RMSE = 0.48), support vector machine regression (SVMR) (R² = 0.78, RMSE = 0.42), random forest (RF) (R² = 0.78, RMSE = 0.42), and extremely randomized trees regression (ETR) (R² = 0.81, RMSE = 0.38). The prediction results were compared with the results of atomic fluorescence spectrometry methods. In the prediction results of the models, the accuracy of ETR using FD preprocessing was the highest. The results confirmed that hyperspectral imaging combined with Spearman’s rank correlation with machine learning models can be used to estimate soil TAs content.