SENet优化的Deeplabv3+滑坡识别

为降低山区滑坡等地质灾害影响的风险和对沿线电网安全的监测保护，利用人工智能实现电力线通道地质灾害的快速筛查。实验利用Deeplabv3+网络对川西地区遥感影像中的滑坡区域进行分割，选取了不同的主干网，包括Resnet、Xception、Mobilenet和引入SE(squeeze-and-excitation)注意力机制的ResNet,实验数据表明，SENet(squeeze-and-excitation networks)优化的Deeplabv3+语义分割模型效果最好，像素准确率(pixel accuracy, PA)达95.5%,平均交并比(mean inetersection over union, MIoU)达84.7%,相比于其他主干网络模型PA最低提升1.3%,MIoU最低提升2.2%。结果表明SENet优化的网络模型在滑坡边缘细节上的分割更精确，误识别和漏识别现象更少，识别效果优于其他模型。

SENet-optimized Deeplabv3+ landslide detection in the Sichuan section of the Sichuan-Tibet Railway

In order to reduce the risk of the Sichuan section of the Sichuan-Tibet Railway being affected by geological disasters such as mountain landslides and to monitor and protect the safety of the power grid along the line, artificial intelligence is used to achieve rapid screening of geological disasters in power line channels. In this experiment, the landslide area in the remote sensing image of western Sichuan was segmented by Deeplabv3+ network, and selects different backbone networks, including Resnet, Xception, Mobilenet, and ResNet (SENet) that introduces the SE attention mechanism. The experimental data shows that the Deeplabv3+ semantics optimized by SENet The segmentation model has the best effect, with a PA of 95.5% and MIoU of 84.7%. Compared with other backbone network models, PA has a minimum increase of 1.3% and MIoU has a minimum increase of 2.2%. The results show that the SEnet optimized network model has more accurate segmentation on the details of the landslide edge, less misrecognition and missed recognition, and the recognition effect is better than other models.