Abstract: Early disaster warning and safety assessment in transmission line corridors are among the priorities of smart grid construction. Therefore, it is very important to grasp the location and distribution of settlements in the transmission line corridor area to do a good job of disaster prevention and mitigation in mountainous areas. In recent years, with the continuous development of target detection technology, the fields it is applied to are becoming more and more extensive, and remote sensing target detection, as one of the application scenarios, is widely used in building information because of its wide coverage and the characteristics of covering many targets. The existing deep learning models have limitations regarding both recognition accuracy and detection speed in building identification and segmentation. Aiming at such problems, this study takes the GF-2 image as the database, labels the buildings in mountainous areas, establishes the sample dataset and divides it into the training set and the test set according to the ratio of 9∶1. Secondly, the standard version of YOLOv7 network is improved by adding a dual-attention module with GAM-CBAM synthesis in the neck part to reduce the feature loss of buildings, which improves the detection ability of the network. The results show that the improved YOLOv7 network achieves an average precision of 88.74% for segmentation and recognition of buildings in mountainous areas, which is also higher than other deep learning models in terms of precision and recall. Therefore, this method can provide data support for rapid and efficient acquisition of mountainous area settlement information, geographic information analysis in the process of disaster prevention and mitigation in mountainous areas, and the development of emergency plans.

Key words: building recognition, machine learning, YOLOv7, transmission line corridor, target detection