Enhancing Lightweight SAR Image Classification Using MobileNetV3-Small Implemented with Convolutional Block Attention Module (CBAM)

Abstract

The rapid worldwide advancement of environmental remote sensing technologies has created a growing need for automated, efficient, and high-performing analysis methods for Synthetic Aperture Radar (SAR) imagery. SAR is particularly valuable because it provides consistent imaging capabilities regardless of weather or lighting conditions, making it ideal for crucial applications such as environmental monitoring and disaster management. Despite these advantages, SAR image classification remains a challenging task due to the complex nature of SAR data, which often includes speckle noise, limited textural information, and high computational demands. To address these challenges, this study introduces an efficient and lightweight Convolutional Neural Network (CNN) architecture based on MobileNetV3-Small, integrated with a Convolutional Block Attention Module (CBAM). The attention mechanism enhances the extraction of meaningful features while reducing unwanted backscatter effects in SAR feature maps. The proposed hybrid model was trained and evaluated on a Sentinel-1 SAR dataset comprising four land cover categories: Agriculture, Barren Land, Grassland, and Urban areas. Implementing data augmentation and transfer learning techniques, the model was trained over 50 epochs, and achieved the high accuracy of 95.74%. This results, outperforming the original MobileNetV3-Small backbone, stems directly from the implementation of CBAM attention mechanism, allowing the proposed model to successfully emphasize relevant spatial and channel features. In this work, it was demonstrated that the proposed MobileNetV3-Small with CBAM architecture delivers a powerful balance between computational efficiency, robust accuracy, and generalization, capability, making it a suitable for integration in near-real-time applications and embedded systems. The results prove the potential of lightweight attention- augmented CNN architecture for a generation of autonomous and intelligent Earth observation systems, offering opportunities for further research in global ecological monitoring.