Multi-Scale Compact Convolution and Attention-Guided Feature Augmentation: An Efficient Lightweight Network for Real-Time Satellite Image Classification

Abstract

A model with high recognition accuracy and minimal processing is required for real-time satellite picture classification. the goal of this research is to create a lightweight Convolutional Neural Network architecture for satellite image classification that satisfies the demands of quick inference speed and high classification accuracy. To learn land-cover structures at various sizes, a multi-scale compact convolution module is added after an effective convolutional backbone is constructed using factorized convolution. The discriminative area is strengthened and redundant background responses are further suppressed by the addition of an attention-guided feature augmentation unit. The experiment uses a collection of 2100 photos from the 21-class satellite scene, with an input size of 224 × 224 × 3. The suggested model has roughly 1.35 million parameters and 185 MFLOPs in size, with a macro-F1 score of 94.31% and a total accuracy of 94.76%. The model has a comparatively low inference cost and performs 1.43% to 5.24% better in classification accuracy than representative lightweight networks. The findings indicate that, in addition to decreasing network size, maintaining multi-scale spatial information and selective feature discrimination is required to enhance the effectiveness of lightweight satellite image classification. The suggested structure is appropriate for onboard remote sensing and real-time satellite picture classification at the edge.