A Deep Learning-Based Strategy for Characterizing the Condition of Roadside Ditches in Infrastructure

The characterization of roadside drainage ditches in infrastructure is essential for road preservation, particularly in regions with high rainfall and mountainous terrain. This study proposes a Deep Learning-based strategy that combines instance segmentation with YOLOv11-seg and classification with EfficientNet-B4 to enhance the detection and analysis of ditches. The strategy follows the CRISP-DM standard, focusing on data understanding and modeling. The training of YOL011vn-seg and YOLOv11x-seg demonstrates high segmentation accuracy, with mAP50 scores exceeding 85%, while classification using EfficientNet-B4 achieves an average accuracy of 99%, minimizing false positives. The results validate that the proposed strategy reduces segmentation errors and improves ditch identification compared to structurally similar objects. It is concluded that the combination of segmentation and classification enables an efficient characterization process, ensuring its applicability in road maintenance.