Identification of Coal Mine Locations using Deep Learning Techniques

Coal is an important energy source for both the local and global economy. Coal mines are also a source of CH4 emissions, the second most important greenhouse gas. Monitoring CH4 emissions caused by coal mining, from space, requires exact location of coal mines. Identifying coal mines can be treated as an image classification problem wherein pixels are clustered into different Land Use/Land Cover categories based on specific features and spectral characteristics. However, open cast coal mines are composed of different features like barren land, industrial complex, storage areas, transport facilities etc. and their spectral characteristics may vary from country to country. Convolutional Neural Networks (CNN) have proved to be capable of such complex land use/ land cover classification tasks. With a list of known coal mine locations, a dataset of coal mines from different countries was prepared using the Sentinel-2 satellite images with 13 spectral bands. Following the EuroSAT dataset as a benchmark, multiple image patches were created for each coal mine location. A total of 3500 coal mine image patches along with image patches of other land use/ land cover features are used to train three deep learning models (VGG, ResNet, DenseNet) with different network architectures. Since all three architectures were similar in performance the simplest architecture (VGG) using small convolutional filters combined with transfer learning was chosen to reduce the computational time. An overall classification accuracy of 98% has been achieved. The model was tested on images outside the training dataset from different countries and found to perform well in identifying coal mines.