Application and Evaluation of Deep Neural Networks for Airborne Hyperspectral Remote Sensing Mineral Mapping: A Case Study of the Baiyanghe Uranium Deposit in Northwestern Xinjiang, China

Deep learning is a popular topic in machine and artificial intelligence research has achieved remarkable results various fields. In geological remote sensing, mineral mapping an appealing application of hyperspectral sensing for surveyors. Whether deep can improve the identification ability images, especially discrimination spectrally similar intimately mixed minerals, needs to be evaluated. this study, shortwave airborne spectrographic imager (SASI) images Baiyanghe uranium deposit Northwestern Xinjiang, China, were used as experimental data. Three neural network (DNN) models designed: fully connected (FCNN), one-dimensional convolutional (1D CNN), two-dimensional 2D CNN). A sample dataset containing five minerals was constructed model training validation, which divided into training, validation test sets at ratio 6:2:2. The final accuracies FCNN, 1D CNN, CNN 91.24%, 93.67% 94.77%, respectively. three DNNs SASI mining area. compared with support vector (SVM) mixture-tuned matched filtering (MTMF) method. Combined ground spectral data obtained by spectrometer, verification interpretation carried out on sections that two kinds methods identified differently. show more accurate than those other methods. More importantly, characteristics, such short-wavelength white mica medium-wavelength mica, had effect DNN models, indicating introduction spatial information images. general, CNNs have good prospects are worthy further development future work.