SHRS-YOLO: Small Target Detection in Hazy Remote Sensing Images

Small target detection in remote sensing (RS) images under hazy conditions faces significant challenges due to atmospheric scattering and absorption phenomena that markedly reduce image contrast and color saturation. Small targets exhibit fragile feature representations, while haze-induced interference complicates target-background discrimination. To address these limitations, we propose SHRS-YOLO, an end-to-end framework specifically designed for small target detection in hazy RS images. The image dehazing (IDH) module incorporates four coordinated filters: dehazing, low-light enhancement, color compensation, and sharpening. A hyperparameter prediction convolutional neural network (HP-CNN) dynamically determines optimal filter parameters based on input image characteristics. The deep feature fusion (DFF) module uses transmission guidance maps to adaptively fuse multi-scale regional attention features with local CNN features, thereby enhancing small target detection capability. Additionally, we introduce an small target detection intersection over union (IoU) loss function with size-based adjustment factors to improve small target detection accuracy. To validate our approach, we synthesized the HazyRS dataset containing 20,000 annotated images with non-uniform haze patterns derived from real satellite imagery. Experimental results demonstrate that SHRS-YOLO achieves 88.37%mAPval50on the HazyRS dataset, representing an 8.40% improvement over baseline methods and a 2.83% improvement over current state-of-the-art (SOTA) approaches.