A Deep Neural Network for Classification of Land Use Satellite Datasets in Mining Environments

Land use (LU) is one of the most imperative pieces of cartographic information used for monitoring the mining environment. The extraction of land use data sets from remotely sensed satellite images has garnered significant interest in the mining region community. However, classification of LUs from satellite images remains a tedious task due to the lack of availability of efficient coal mining related datasets. Deep learning methods provide great leverage to extract meaningful information from high-resolution satellite images. Moreover, the performance of a deep learning classification approach significantly depends on the quality of the datasets. The present work attempts to demonstrate the generation of satellite-based datasets for the performance analysis of different deep neural network (DNN)-based learning algorithms in the LU classifications of mining regions. The mining regions are broadly classified into distinct regions based on visual inspection, namely barren land, built-up areas, waterbody, vegetation, and active coal mines. In our experimental work, a patch of 100 spatial samples for each of the five features is generated on three scales, as [1 × 1 × 3], [5 × 5 × 3], and [10 × 10 × 3]. Moreover, the effects of different scalabilities of the dataset on classification performances are also analyzed. Furthermore, this case study is implemented for the large-scale benchmark of satellite image datasets for mining regions. In the future, this work can be used to classify LU in the relevant study regions in real time.