Natural soil samples are the primary data source of heavy metal hyperspectral predicting models. However, their spectra are often confounded by complex components, resulting in poor performance model and unclear explanation for heavy metal spectral response mechanism. Near standard soil samples provide a promising method for the mechanism research. In this paper, 86 natural soils samples and relatively clean background soil were collected in a lead-zinc mine in Hunan province, and 40 near standard samples were made by artificially adding heavy metals in background soil using control variable method. Feature bands for soil Pb spectra were first selected based on near standard soil samples using partial least squares regressions(PLSR). The feature bands were used to calibrate the prediction model with PLSR for natural soil samples. The existence of Pb absorption features was confirmed by the overall consistent and change trend of near standard samples reflectance spectra. Near standard samples provided acceptable estimation accuracies of Pb concentrations, with the determination coefficient (R2p), and the ratio of prediction to deviation (RPD) values of 0.85 and 2.30. When compared with the entire-band PLSR model, feature-band model for natural soil samples increased the R2p and the RPD from 0.32 and 0.20 to 1.55 and 1.44 by removing uninformative spectral variables. The mechanism investigation strategy we proposed could effectively solve the problem of complex sample composition and weak heavy metal spectral signal in previous research, and be applied in further soil heavy metal remote sensing monitoring.