Comparative Analysis of Pre-trained Architectures for Remote Sensing Image Scene Classification

Remote sensing offers a comprehensive perspective and a practical overview of a region. Through remote sensing, we can monitor the physical characteristics of an area by measuring the radiation reflected and emitted remotely. In most cases, this is done by satellite or aircraft (It can be a drone image). It is possible to interpret a lot of information using Remote Sensing (RS). There is still a challenge in selecting and combining appropriate features from remote sensing imagery based on their spectral and spatial characteristics. A deep artificial neural network is a computing system that has the capacity to learn from data on its own using algorithms. In scene classification approaches, convolutional neural networks have been successfully applied to create a system capable of handling classification without human intervention. The pre-trained classification technique is suggested in this work to categorize the satellite images. The classifiers are implemented for pre-trained deep learning networks including AlexNet, Resnet50, GoogleNet, Vgg16, and Inception V3. The performance is measured in terms of classification accuracy (AC), sensitivity (SE), specificity (SP), and positive predictive value (PPV).