Hyperspectral imaging and deep learning techniques provide new methods and tools for detecting soil contaminants. This study explores a convolutional neural network (CNN)-based algorithm for the detection of hyperspectral soil contaminants. First, a hyperspectral soil dataset containing multiple spectral bands was collected, and data analysis and feature extraction were performed. Subsequently, a CNN architecture adapted to the characteristics of hyperspectral soil data was designed. A self-attention mechanism was introduced to automatically reduce the dimensionality of redundant spectral data, and a feature fusion structure using graph features was employed for feature extraction. Finally, the performance of the algorithm was validated using a collected soil contaminant dataset. The experimental results demonstrate that the proposed method effectively reduces the dimensionality of hyperspectral data, decreases data redundancy, and achieves good performance and accuracy in soil contaminant detection by incorporating graph features. This method is of practical significance for the rapid detection of soil contaminants.