UAV-based automated earthwork progress monitoring using deep learning with image inpainting

Accurate monitoring of earthwork progress is crucial in construction due to its significant costs and potential delays. Traditional methods are labor-intensive and pose safety risks. Unmanned Aerial Vehicle (UAV) photogrammetry offers a promising alternative. However, the presence of moving machinery can distort earthwork progress in generated point clouds. This paper addresses this challenge by integrating deep learning-based segmentation and image inpainting techniques to remove machinery from aerial images. The AIDCON dataset was used to train the Pointrend algorithm for machinery segmentation, achieving an average precision exceeding 90% across common machinery categories. The identified machinery was removed using the LaMa inpainting algorithm. Field experiments validated that the UAV-based net volume calculations closely matched the laser scanning results with less than 6% deviation, and both methods aligned with truck count estimates. Furthermore, the required time was reduced from several days to hours, lowering labor costs and enhancing safety.