AAD-YOLO: An Improved YOLOv8 Model for Complex Remote Sensing Scenarios

Detecting small objects in remote sensing images is challenging due to complex background interference, the tendency for them to be misse, and multi-scale variations. These challenges lead to background noise, blurred boundaries, and loss of edge information, which complicates detection. To address these challenges, we propose the Adaptive Architecture for Detection with Edge Awareness and Dynamic Upsampling (AAD-YOLO). Inspired by the Sobel operator, we designed the C2f-based Adaptive Instance Enhancement Block (CAIEB) to improve the fusion of spatial and edge information. Additionally, the Attention-based Intrascale Feature Interaction (AIFI) module, which incorporates a multi-head self-attention mechanism, focuses on salient object regions while suppressing background interference. The Dynamic Upsampler (Dysample) module enhances multi-scale detection robustness by adaptively adjusting sampling point positions. Experiments on the SIMD dataset demonstrate that AAD-YOLO outperforms YOLOv8n in terms of precision (P), recall (R), and mean average precision (mAP50) by 3.6%, 0.8%, 2.5%, and 1.9%, respectively. On the NWPU VHR-10 dataset, AAD-YOLO achieves a precision of, recall, and average accuracy of 96.3%, 93.5%, and 97.6%. A comparative analysis with several state-of-the-art detection algorithms further validates the effectiveness and robustness of the proposed method.