Design of an optimized deep learning algorithm for automatic classification of high-resolution satellite dataset (LISS IV) for studying land-use patterns in a mining region

Abstract

Accurate classification of the satellite image is always a challenging task, and hence uncertainty involves in the change detection analysis in the land use for mining regions. The present study attempts to demonstrate the development of an optimized deep convolutional neural network with stochastic gradient descent momentum (DCNN + SGDM) using Linear imaging self-scanning sensor- IV (LISS-IV) data for automatic classification of satellites image to assess the land-use pattern in a mining region. An image dataset was prepared by extracting the LISS-IV image samples of five common land-use types (coal mining region, built-up area, barren land, vegetation, and waterbody). A total of 6250 image samples (1250 for each class) of 6 × 6 patch size were generated for the model training and testing. The study area selected for analyzing the model performance and impact analysis is Jharia coalfield (JCF), Jharkhand. The study area consists of 5184 × 7794 pixels. The model performance results indicated that classification accuracies on the training and testing dataset are 99.04% and 83.50%, respectively. The model performance was also compared to the performance reported in the well-established deep learning algorithms, including Alex Net, Squeeze Net, Google Net, VGG Net 16, and VGG Net 19. It was observed that the DCNN + SGDM model performs better (99.04%) than the existing deep learning models (73.8% in Alex Net, 80.3 in Squeeze Net, 93.33 in Google Net, 93.20% in VGG Net 16, and 93.00 in VGG Net 19). Moreover, the DCNN + SGDM model's performance was compared with the two pixel-based classification algorithms (support vector machine (SVM) and deep neural network (DNN)) in terms of accuracy and processing time. The comparative values of accuracies are 70.40% for SVM, 69.10% for DNN, 99.04% for DCNN on the training dataset, and the respective elapsed time of the SVM, DNN, and DCNN models are 1.716 s, 3.59 s, and 1587 s. That is, the processing time of the DCNN model is high but offers higher accuracy in comparison to the pixel-based classification algorithms (SVM and DNN).