This paper aims to explore the optimal inversion model of regional heavy metal content in soil. With Longhai City taken as the study area, this study preprocessed the original spectral data of soil using the methods of Savizky Golay (SG), wavelet transform (WT), gaussian filter (GF), and multiple scatter correction (MSC) individually, then eliminated the interference and wavelength bearing no information using the wavelength selection algorithms developed based on model population analysis (MPA), including the competitive adaptive reweighted sampling (CARS), variable iterative space shrinkage approach (VISSA), iteratively variable subset optimization (IVSO), and interval combination optimization (ICO), and finally predicted the lead content in soil using the linear partial least squares regression (PLSR) model, nonlinear support vector machine (SVM) model, and extreme learning machine (ELM) based on neural network. The results are as follows. ① Among the inversion models of lead content in soil established using various preprocessing methods, the model built based on reconstructed spectral data of level 7th by wavelet transform had the most optimal prediction accuracy, with R²=0.736, RMSE=5.426, RPD=1.976, and RPIQ=2.560. ② The CARS, VISSA, IVSO, and ICO algorithms developed based on MPA significantly improved the performance of model interpretation and generalization and improved modeling efficiency. ③ In terms of overall prediction results, the three regression models were in the order of ELM>PLSR>SVM. Among them, the ICO-ELM had the highest prediction accuracy, with R²=0.863, RMSE=3.953, RPD=2.712,and RPIQ=3.514. Therefore, the optimal model established in this study can provide a new theoretical reference for the rapid monitoring of regional land quality and ecological indicators.