Taking 80 black soil samples collected from Nehe City, Heilongjiang Province, and hyperspectral measured data as data sources, the spectral reflectance and its feature changes of copper (Cu), zinc (Zn), manganese (Mn) in black soil were analyzed, the correlations between four different forms of spectral reflectance, which included original, first-order differential, continuum removal, and firstorder derivative of continuum removal and soil Cu, Zn, Mn contents were calculated, and the correlation coefficient method was used to extract sensitive bands. Then the kernel principal component analysis (KPCA) was applied for dimension reduction and feature extraction of hyperspectral sensitive band data, and the feature information was input into extreme learning machine (ELM), and the KPCA-ELM estimation model was constructed to quantitatively estimate the heavy metal contents. The results showed that KPCA had a strong ability to extract nonlinear features and effectively selected the optimal variable set. The KPCA-ELM model was feasible in predicting soil element content and the determination coefficients of the three heavy metal elements were all more than 0.6, where the prediction accuracy of Zn was the highest among the three heavy metal elements. And the determination coefficient and root mean square error were 0.805 and 3.275mg/kg respectively, which were improved by 14.0% and 18.5% compared with without feature extraction. Therefore, KPCA-ELM model was a fast and feasible method for hyperspectral estimation of heavy metal content.