Abstract: In contemporary hydraulic engineering management, daily inspections and specialized surveys generate numerous on-site images. However, the images captured by multiple teams are often uploaded without classification, leading to information overload and ineffective use of the historical images for data analysis. To overcome the low efficiency of manual classification, this paper describes an intelligent archiving framework based on an improved YOLOv8 detector, enabling automated processing from object detection to image classification. We construct a dedicated dataset equipped with 12 types of typical hydraulic engineering targets, integrate the EMA attention mechanism to enhance multi-scale feature representation, and adopt the SIoU loss function to raise the bounding box regression accuracy. Experimental results show this new archiving model achieves mAP50-95 of 0.624 and mAP50 of 0.793, or an improvements of 9.00% and 8.83% over the baseline model, respectively. It leverages the structured information－such as target categories, positions, and scales, derived from detection outputs－and therefore enables the multi-dimensional automatic archiving of images according to the engineering components and work scenarios, achieving an accuracy of 95.4% on a test set comprising 3,348 images. The detection-driven archiving (DDA) method demonstrates an effective approach to the intelligent management of hydraulic engineering image datasets.

Key words: hydraulic engineering images, object detection, machine vision, YOLOv8, intelligent archiving, attention mechanism, loss function