AGW-YOLO-Based UAV Remote Sensing Approach for Monitoring Levee Cracks

ObjectiveRecently, the identification of cracks in river levees primarily relies on manual inspections, which suffer from low efficiency, high costs, and significant safety risks. Although deep learning technologies have made notable progress in enhancing detection automation, several challenges remain, including insufficient detection accuracy, poor adaptability, and a lack of precise localization capabilities. To address these issues, this study proposes a lightweight real-time levee crack detection and localization method—AGW-YOLO, utilizing UAV remote sensing imagery for crack identification. The method aims to ensure high detection accuracy with less computing cost. It enables real-time recognition of the geographic coordinates of cracks, thus providing technical support for the safety monitoring and subsequent maintenance of water conservancy projects.MethodsTo address challenges such as the difficulty in capturing subtle features and the limitations in computational resources during UAV-based levee crack inspection tasks, the AGW-YOLO object detection model was introduced. Firstly, a lightweight ADown module was incorporated to replace the conventional stride-2 convolution. The ADown module dynamically adapts its downsampling strategy according to the feature characteristics, effectively reducing the number of parameters and computational complexity, while enhancing the model's ability to capture crack edges and fine textural details. Secondly, a novel Global Multi-Scale Attention (GMA) mechanism was proposed to improve the recognition of both local and global crack features. This mechanism groups the input feature maps by channels and applies average and max pooling operations in both the horizontal and vertical directions. Attention weights are then generated to simultaneously fuse spatial and channel information, significantly improving the model's accuracy in identifying crack regions in levees, particularly under complex lighting and background interference. Finally, to address the multi-scale variation of cracks, the original loss function was replaced with WIoUv3. This loss function incorporates an outlier degree evaluation and a dynamic non-monotonic focusing mechanism, which adaptively adjusts the optimization intensity of anchor boxes with varying qualities. This approach effectively prevents overemphasis on high-quality anchors and underemphasis on low-quality ones during training, ensuring a balanced regression accuracy across all bounding boxes while minimizing computational burden. For crack localization, a high-precision method based on UAV orthophotos was proposed by combining the crack detection results with the GPS coordinate information. Initially, the WGS84 coordinates of the image center were transformed into UTM coordinates. Subsequently, the center pixel coordinates of the bounding boxes predicted by the detection model were integrated with the yaw angle data recorded in the UAV's flight logs to construct an Euler angle rotation matrix. This matrix was then employed to compute the actual crack positions within the UTM coordinate system. Finally, the calculated coordinates were converted back to the WGS84 system, enabling precise mapping from pixel coordinates to geographic coordinates.Results and Discussions To validate the effectiveness of the proposed method, the "levee crack" dataset for levee crack detection was built. Data collection was conducted using the DJI Matrice 300 RTK drone platform, with multiple systematic aerial photography missions performed over the target levee regions during 2022 and 2023. The dataset focused on three typical crack categories: ordinary cracks, repaired cracks, and alligator cracks. To enhance the dataset's generalization performance, various data augmentation techniques, including random image rotation, were applied. Ablation and comparative experiments were carried out to assess the performance of the AGW-YOLO model in terms of detection accuracy and real-time capability. Ablation results revealed that incorporating the ADown module reduced the baseline's parameters by 18.6%, although it led to a slight decrease in recognition accuracy and recall. Using the GMA attention mechanism, the model demonstrated enhanced precision in extracting crack edges and texture details, with a minimal increase of only 0.48% in the parameters. This modification resulted in significant improvements in recall rate and overall mean Average Precision (mAP). The advanced WIoUv3 loss function further boosted the model's performance, achieving a [email protected] of 84.5% and an F1 score of 83%, marking an approximate 3.4% improvement over the baseline, and showcasing a favorable balance between detection accuracy and model efficiency. Comparative experiments indicated that AGW-YOLO outperformed several mainstream object detection algorithms, including Faster R-CNN, YOLOv8n, YOLOv9-tiny, YOLOv10n, RT-DETR-R50, and TPH-YOLO, across most evaluation metrics, offering high recognition accuracy with lower computational complexity. Furthermore, crack localization experiments demonstrated that the proposed method, based on UAV orthophotos, achieved root mean square errors of 1.753 meters and 1.053 meters in the east and north directions, respectively, within the UTM coordinate system. The maximum error was confined to within ±3 meters, meeting the spatial localization accuracy requirements for large-scale water conservancy projects.ConclusionsThis study addressed critical challenges in the identification and localization of cracks in levees for water conservancy projects. A novel crack detection and localization approach, utilizing UAV remote sensing imagery in conjunction with a lightweight AGW-YOLO model, was proposed and demonstrated excellent performance in practical applications. The incorporation of the Adown module, GMA attention mechanism, and WIoUv3 loss function effectively reduced the model's parameters and computational complexity, while significantly improving the accuracy and robustness of levee crack detection. Furthermore, the integration of crack detection results with coordinate transformation algorithms enabled the precise geospatial localization of cracks. This approach met the stringent accuracy requirements for identifying both cracks and potential hazards in water conservancy infrastructures and provided an efficient and practical solution for levee safety monitoring. Future research may focus on combining AGW-YOLO with UAV swarm systems to enable fully automated, scheduled, and intelligent crack detection and analysis of levees, thereby offering robust technical support for long-term safety monitoring in hydraulic engineering.