A lightweight remote sensing image detection model with feature aggregation diffusion network

Abstract

With accelerating land-use changes driven by urbanization and resource extraction, accurate detection of landscape objects in remote sensing imagery has become pivotal for sustainable land management. However, existing deep learning models often face challenges in balancing detection accuracy and computational efficiency, especially for small objects in complex scenes. To address this, we propose LightFAD-YOLO, a lightweight model integrating feature aggregation diffusion for multi-scale context propagation, enhancing small object detection in complex scenes. The central convolutional detection head combines detail-enhanced convolution and group normalization, reducing computational costs by 23.4 % while maintaining precision. A dilation-wise residual module further optimizes multi-scale feature extraction. Evaluated on benchmark datasets, LightFAD-YOLO achieves 1.7 % higher ${mAP}_{0.5}$ and 6.4 % improved ${mAP}_{0.5:0.95}$ over baseline models, with 9.9 % lower computational load. Operating at 297.2 FPS with only 2.3M parameters, it enables real-time deployment on edge devices for land-use monitoring and infrastructure detection, supporting sustainable land management.