Band Selection Method for Retrieving Soil Lead Content with Hyperspectral Remote Sensing Data

Hyperspectral data offers a powerful tool for predicting soil heavy metal contamination due to its high spectral resolution and many continuous bands. However, band selection is the prerequisite to accurately invert and predict soil heavy metal concentration by hyperspectral data. In this paper, 181 soil samples were collected from the suburb of Nanjing City, and their reflectance spectra and soil lead concentrations were measured in the laboratory. Based on these dataset, we compare Least Angle Regression, which is a modest forward choose method, and least squares regression and partial least squares regression based on genetic algorithm. As a result, regression with band selection has better accuracy than those without band selection. Although both Least Angle Regression and partial least squares regression with genetic algorithm can reach 70% training accuracy, the latter based on genetic algorithm is better, because it can reach a larger solution space. At last, we conclude that partial least squares regression is a good choice for the soil lead content retrieval by hyperspectral remote sensing data, and genetic algorithm can improve the retrieval by band selection promisingly. Bands centered around 838nm,1930nm and 2148nm are sensitive for soil lead content.