Categorization and Interpretation of Satellite Image Scenes Employing AI Approaches

Scene identification in Very High-Resolution (VHR) photography presents a formidable challenge. Although Convolutional Neural Networks (CNNs) have enhanced accuracy in feature learning, their deep layers often struggle to accurately depict object relationships within images. To address this limitation, the paper introduces an advanced Multilayer Perceptron (MLP) acting as a deep classifier, utilizing RMSprop and Adadelta optimizers for classification. Our proposed model, CNN-MLP, merges the strengths of MLP and CNN methods. It utilizes a pre-trained CNN, devoid of fully-connected layers, for feature generation, supplemented by data augmentation (DA) techniques to enrich the training dataset. The resulting feature maps undergo classification using an MLP, achieving an outstanding classification performance. The model excels in identifying barren and farm land, even within the same image, showcasing its efficacy in scene classification. This success is demonstrated using three publicly available VHR image datasets UC-Merced, Aerial Image (AID), RSI CB 128, NWPURESISC45 combined to and also create a blended dataset with the overall 96.5