A deep residual convolutional neural network for mineral classification

In recent years, the deep learning computing paradigm has revolutionized the way of remote sensing data analysis. The emerging hyperspectral remote sensing has paved the way for the efficient and accurate exploration of the minute features of the earth’s surface due to the increased number of contiguous spectral bands. Hence, hyperspectral images can efficiently impart useful information about mineral resources for precise discrimination and identification in lithological studies. Rapid advancements in computing capabilities and deep learning techniques give the research community a new impulse to develop advanced, robust, and efficient hyperspectral remote sensing-based mineral classification frameworks. The present study aims to introduce two novel deep learning-based mineral classification frameworks: mineral-CNN-LSTM and mineral-ResNet. The architecture of mineral-CNN-LSTM is based on 1D-CNN and LSTM model, whereas the architecture of mineral-ResNet is based on 1D-CNN, LSTM model, and residual connections. The frameworks use raw data as input without feature selection or data augmentation preprocessing steps. The widely used early stop method is also utilized to prevent overfitting of the framework during the training process. The experimental evaluation carried out over the AVIRIS hyperspectral image scene of the Cuprite mining area confirms that the mineral-ResNet can effectively identify most of the minerals such as Alunite, Calcite, Halloysite, Kaolinite, Montmorillonite, Muscovite, Chalcedony with the overall accuracy of 92.16%, and kappa value of 0.89 and mineral-CNN-LSTM achieved the overall accuracy of 91.71% and kappa value of 0.88 for these minerals. Furthermore, a comparative evaluation of the proposed frameworks has been performed with widely used Convolutional Neural Network (CNN) based architectures such as VGG19, VGG16, ResNet-50, and AlexNet; and various machine learning based classifiers. The proposed architectures offer better performance with shorter testing and training time than these existing CNN-based architectures. The proposed framework could be useful for other earth observation-related applications in various fields such as agriculture, forestry, geology, hydrology, ecology, urban planning, military and defense applications, etc.