Pothole detection and risk zone analysis using YOLOv9 with adaptive unsharp mask-guided filter

Pothole is the major type of structural defect found on roads, commonly generated due to thin or weak structure. However, existing models still struggle with low-contrast images, irregular pothole shapes and complex road textures that reduce detection accuracy. Furthermore, you only look once (YOLO) lacks the ability to interpret spatial risk zones that are necessary for assessing the severity of road damage. To overcome these challenges, a novel modified YOLOv9 is proposed that integrates an Adaptive Unsharp Mask-Guided filter and Programmable Gradient Information (PGI)–GELAN feature aggregation to achieve accurate pothole detection and risk-zone analysis. The proposed modified YOLOv9 is evaluated on the public pothole dataset, which includes potholes of varying shapes, shadows and illumination conditions. It effectively identifies the pothole area, including coverage, flow direction, weak zones and boundaries. The proposed modified YOLOv9 model achieves the total accuracy of 99.41% based on the gathered dataset. Comparing the modified YOLOv9 model shows advances in FPS of 16.5% and increases in mAP of 4.23%. An ablation study shows that accuracy drops to 95.71% without AUM, 96.35% without PGI and 97.27% without the GELAN block while integrating all modules enables the modified YOLOv9 to achieve the highest accuracy of 99.41% for pothole detection.