Revolutionizing Satellite Image Classification using Deep Learning Framework

The use of satellite imagery to classify scenes is of high importance in environment monitoring, surveillance, and in analyzing the scenes with other intentions for different functions. The recent developments of Remote Sensing technologies have helped in the extraction of spatiotemporal data. Several techniques are advanced to perform automatic classification of the land scenes of remote sensing images. However, these methods are confronted with some issues when attempting to identify various scenes on the land. Heavy occlusion, complex background, objects which are too small to be recognized easily and objects of highly variable sizes are the issues which limit the models to a good degree of accuracy. In order to overcome these challenges a ResNet-50 based deep learning framework is put forward to extract global features to effectively detect the complicated land scenes. In general, this framework is divided into two limbs. There the first branch deals with preprocessed grayscale image from different scenes while the next one investigates the so-called CLAHE preprocessed features. Performance analysis of the proposed framework is carried out on two standard databases Aerial image dataset and EuroSAT. The results of empirical observations based on ResNet-50 gives the highest accuracy of the analysis for all calculated criteria, including accuracy rate, sensibility, and the specificity level compared to other similar models.