<![CDATA[In the past two decades, remote sensing-based landslide detection methods have advanced significantly. However, these methods generally remain broad in scope and are not yet capable of identifying landslides as distinct objects with precise locations. This study aims to integrate YOLO v3 into remote sensing applications to enhance the accuracy of landslide detection in sub-watershed (sub-DAS) areas. The study focuses on two sub-watersheds with a history of frequent landslides, namely Kali Konto and Sumber Brantas. Sentinel-2A imagery from 2024 was downloaded for all months and composited to achieve the most stable visualization. Landslide location points were collected through an exploratory approach, incorporating information from the National Disaster Management Agency (BNPB) and local communities. A total of 155 landslide points were recorded in the field and subsequently used as training and validation data. Seventy percent of these points were analyzed using a deep convolutional neural network—YOLO v3—implemented in ArcGIS Pro, while the remaining 30% were used for validation against actual field occurrences.The detection model developed for the Kali Konto and Sumber Brantas sub-watersheds achieved an accuracy of 77%, demonstrating reliable performance in predicting landslide locations. The landslides identified predominantly consisted of slope failures and rockfalls occurring along cut-fill sections of roadside areas. However, this study has certain limitations: the developed model is unable to classify different types of landslides, and the satellite imagery used has a medium resolution, which is not yet advanced enough to fully meet the requirements for sub-watershed-level analyses.]]>
Copyrights © 2026
