Inversion Study of Heavy Metals in Soils of Potentially Polluted Sites Based on UAV Hyperspectral Data and Machine Learning Algorithms

We established a risk screening and grading model and a content estimation model for zinc pollution in bare soil. We built these models using the machine learning algorithms Support Vector Machine (SVM), Generalized Linear Model (GLM), Multivariate Adaptive Regression Spline (Mars), Random Forest (RF), XGBoost, Ridge Regression (Ridge), and Cubist based on UAV hyperspectral data and heavy metal field fast detection data in typical potentially contaminated sites. The parameters of the hyperspectral data were original reflectivity, smoothed reflectivity, first-order derivative, second-order derivative, and the de-enveloping spectrum. The results showed that to classify soil zinc pollution risk, the machine learning model based on the second-order derivative spectrum performed better than the other hyperspectral parameter independent variables. The overall classification accuracy of the MARS model based on the second-order derivative spectrum was 89.29%. The XGBoost model based on the second-order derivative spectrum performed the best in estimating zinc content, with results of R ² = 0.59. When the zinc (Zn) content was less than 1000mg/kg, the model accuracy was stable. This method doesn't rely on soil samples, and thus avoids uncertainty caused by the selection of sensitive bands in heavy metal inversion. This method provides a basis for large-scale fast investigation of soil heavy metal pollution based on limited ground monitoring point data.