A rapid detection and quantification method for levee leakage outlets using drone infrared thermography and semantic segmentation

Leaking erosion is one of the most harmful factors causing levee breaches. To ensure the safety of levees, rapid leakage detection and evaluation is imperative. In current practice, manual operation-which is risky and inefficient-remains the most adopted method and lacks modern intervention. To bridge this gap, a novel fully automated method utilizing drones, infrared thermography, and deep learning is proposed. Drones equipped with infrared and visible imagers are utilized to perform fast sensing of distributed levees, and the sensing data is communicated to a server where pre-trained models are deployed for leakage exit recognition and quantification. To build such a service system, three different thermal imagers are employed to collect leakage images in a variety of leaking scenarios, including laboratory models and field actual levees. A total of 4686 raw infrared images containing diverse leakage objects are selected out and carefully annotated by hand at the pixel level. Afterwards, 14 models are trained using 5-fold cross-validation, taking into consideration different backbones and class imbalance. The highest performance model achieved a mean intersection over union value of 82.73% on the test set, despite the test set including many challenging samples. To assess the practicality of the proposed approach, a full-scale levee section is constructed. Several field tests demonstrated the effective of the proposed detection and quantification strategy. The well-trained detector demonstrated excellent performance in terms of practicability and generalization. The influencing factors and limitations of the proposed strategy are thoroughly examined.