Enhancing Remote Sensing Scene Classification with Channel-Spatial CNN (CS-CNN)

Classification of remote sensing scenes is crucial for the analysis of the Earth’s surface using satellite or aerial imagery. The classification of remote images unlocks a world of applications, including Land Use Planning, Urban Development, Environmental Monitoring, Disaster Management, Defence, and Geospatial Analysis. The standard Convolutional Neural Networks (CNNs) often struggle to extract class-relevant information from high-quality remote sensing images. In response to this challenge, our research introduces an innovative approach: the integration of a CNN with a channel and spatial attention mechanism, aiming to achieve precise classification of remote scene images. Unlike conventional CNNs, which primarily extract fundamental features, our proposed approach aims to extract both channel-based and spatial-based features from the respective images to achieve improved performance. The proposed Channel-Spatial Convolutional Neural Network (CS-CNN) model automatically categorizes the RSSCN7 dataset images distinct classes, including Grass, Field, Industry, River Lake, Forest, Residential, and Parking lot. This model enhances the accuracy while concurrently diminishing loss in the classification system. It accomplishes this feat by directing its attention to pertinent regions within the images and effectively capturing long-range dependencies inherent in remote sensing imagery. To assess the efficacy of our proposed method, we conducted an extensive evaluation involving multiple deep learning models, encompassing standard CNNs, CNNs enriched with channel and spatial attention mechanisms, ResNet50, ResNet50 with channel attention, VGG19, and VGG19 with attention mechanisms. Experimental results demonstrate the superiority of the CNN integrated with both channel and spatial attention mechanisms, achieving an impressive accuracy rate of 96% in the classification of remote images.