Lightweight deep learning for remote sensing scene classification

Abstract

Remote sensing scene classification plays a vital role in various fields, including environmental monitoring, defence, and surveillance. Recent advances in convolutional neural networks (CNNs) have significantly improved classification accuracy; however, these improvements often come at the expense of increased model size. Although such models achieve high accuracy, they are not well-suited to applications requiring rapid responses, such as disaster management, where timely decision-making is critical. Furthermore, deploying these large models on edge devices, such as drones, which are commonly utilised in disaster response, presents challenges due to limited memory and processing power.

This report examines lightweight architectures to enhance classification efficiency, specifically analysing the EfficientNet-B3 model from the EfficientNetV1 family. Light\-weight models frequently employ attention mechanisms to prioritise essential features during classification. However, many existing channel attention mechanisms rely on global pooling operators, which may result in the loss of crucial spatial information.

To address this issue, two novel channel attention mechanisms are introduced: the Quadrant Attention Module (QAM) and the Hybrid Attention Module (HAM). QAM enhances spatial awareness by capturing localised feature interactions, while HAM further refines this approach by incorporating broader contextual information. Experimental results demonstrate that integrating the HAM module into the EfficientNet-B3 model outperforms existing state-of-the-art channel attention mechanisms, underscoring its effectiveness in remote sensing scene classification. Notably, EfficientNet-B3 HAM exhibits the potential to surpass EfficientNet-B3 integrated with both channel and spatial attention modules. This highlights its ability to capture both channel-wise and spatial information without requiring separate spatial attention mechanisms.