Automated localization of landslide features on historical geological maps using deep learning

Abstract

The existence of small dimensions, thin lines, background noise, and colour changes due to paper wear makes manual symbol localization difficult on old geological maps. A method that addresses longstanding challenges in archival map utilization, such as collating heterogeneous symbols in old cartographic materials, is essential. Therefore, artificial intelligence can be beneficial in this field. By automating symbol localization, it unlocks previously underutilized archives for modern hazard studies, particularly in areas with limited contemporary monitoring. This study uses deep learning models and introduces a novel technique to automate the oriented localization of landslide-related symbols across a geological map series with a scale of 1:25,000 from the archive of the State Office for Geology and Mining Rhineland-Palatinate, Germany. Initially, geological images were cropped into tiles of two sizes: 128 × 128 and 416 × 416 pixels. Then, using established deep learning models, automatic localization of symbols using oriented bounding boxes (OBB) was performed at a single-scale and dual-scale by combining the detections at two scales. The YOLOv8 and v11 OBB models exhibited the best accuracy and detection performance among the deep learning models. Moreover, it was found that the dual-scale detection method achieved higher overall accuracy compared to the single-scale models, although it required significantly higher runtime. Furthermore, the 416 × 416-pixel tile size provided both lower runtime and higher detection accuracy than the 128 × 128-pixel tile, indicating that larger tiles enable more complete object representation and more efficient inference. Finally, a new 4-channel input method was introduced and used in the YOLOv11 OBB architecture. A detailed analysis of 94 historical and contemporary maps spanning 129 years (1887–2016) using this novel technique was able to provide accuracy at least at the same level as the dual-scale model. Therefore, the 4-channel model is considered the most optimal and efficient configuration for map symbol localization.