This work is devoted to the development of a mine detection method based on processing hyperspectral images obtained using a hyperspectral camera installed on board an unmanned aerial vehicle (multicopter). The main problems that arise in this case are identified: high correlation between spectral bands, spatial variability of various spectral features and the curse of dimensionality, the presence of a large amount of unlabeled data. To solve this problem, the work uses the feature extraction (FE) method, which selects key spectral ranges while preserving the physical content and valuable spectral-spatial properties of hyperspectral images (HSI). We study a semi-supervised algorithm for training a convolutionalS neural network using a denoising autoencoder. The basis of this autoencoder is a modified network using 3D convolution and a modified hybrid convolutional network using 2D convolution. This network was trained using the Adam W algorithm, which is a modification of the Adam algorithm with separate regularization of weight decay. The developed algorithms were tested on test data.