YOLOv8‐DBW: An Improved High‐Accuracy Fast Landslide Detection Model

ABSTRACT Landslides can pose a significant threat and damage to life and property. Rapid and accurate determination of the location of landslides is critical to preventing and stopping geohazards. Currently, manual visual interpretation, semiautomatic interpretation, and traditional machine learning interpretation of landslides are less efficient and cannot meet the requirements of certain high‐standard interpretation, while the object detection technology based on deep learning is developing rapidly, and its application in the field of intelligent interpretation of landslides is still immature. In this paper, YOLOv5, YOLOv6, YOLOv7, and YOLOv8 algorithms are evaluated to obtain the ideal landslide detection effect. In the respective variant models of YOLOv5, YOLOv6, and YOLOv7, the corresponding optimal landslide detection variants are YOLOv5x, YOLOv6n, and YOLOv7. Among the variant models of YOLOv8, YOLOv8n requires the least number of parameters and has the fastest detection speed. However, because the YOLOv8n model has the potential to be optimized in three aspects, such as feature fusion, attentional mechanism, and loss function, it was improved in this study in these three aspects. In this study, the improved model was named the YOLOv8‐DBW model. Compared with YOLOv8n, this model has three main advantages: (1) Precision, recall, F1 score, mAP 0.5 , and mAP 0.5:0.95 are improved by 11.8%, 2.0%, 7.3%, 12.4%, and 5.4%, respectively; (2) higher detection accuracy, more accurate localization, and determination of the extent of landslides, as well as lower leakage rate; (3) the model can detect landslides of different sizes completely and accurately when in the scenarios of single landslide, multiple landslides, and multiple disturbances. This model still has potential for optimization in terms of computational resource consumption and detection of generalizability.