In order to capture the complex non-linear relationships in soil spectra, a spectral multi-channel convolution method based on convolutional neural networks was proposed, and the soil heavy metal nickel content was predicted. One hundred and twenty-two soil spectral samples were collected from polluted agricultural soils in an area. The Kennard-Stone algorithm was used to divide the samples into calibration sets and verification sets. The Savitzky-Golay smoothing and the standard normal transformation were used to preprocess the original soil spectral data, and an improved correlation coefficient method ($ p $=0.01) was used to extract 296 characteristic bands. Four deep learning methods, including ResNet, VGG, Inception, and MobileNet, were employed to predict the content of heavy metal nickel under different channel strategies: single-channel (MTF), dual-channel (MTF-GASF), and multi-channel (MC). Without increasing the model complexity, a method was proposed to enhance the prediction accuracy of the lightweight model MC-MobileNet for soil nickel content. In order to comprehensively evaluate the prediction performance of different models, three indicators, including determination coefficient, root mean square error and relative predictive deviation, were used for evaluation. Results showed that the prediction performance of all models was improved after using the multi-channel convolution method, the model overfitting scenario was mitigated, and the models under the multi-channel strategy exhibited a relative predictive deviation greater than 2.5.