Improved Mask R-CNN for Disturbed Area Extraction in Construction Projects from High-Resolution Satellite Imagery

This study introduces an enhanced Mask R-CNN model for automatically extracting disturbed areas in construction projects from satellite imagery, aiding soil and water conservation authorities in monitoring tasks. The model optimizes the feature extraction network structure and loss function to bolster object recognition and segmentation accuracy. By integrating a channel attention mechanism, SENet, the model’s capacity to detect target objects is augmented. The loss function optimization results in more precise edge detection. Utilizing the high-resolution satellite data from GF-1, the effectiveness of the proposed method has been successfully demonstrated in the study area of Liuyang City, Hunan Province, China. The proposed model exhibits substantial improvement over the traditional Mask R-CNN, delivering accurate segmentation and precise identification of disturbed areas in construction projects.