基于单阶段实例分割网络的黄土滑坡多任务自动识别

滑坡自动识别能够解决人工目视解译方法速度慢的问题，现有基于深度学习的自动识别方法多以目标检测和语义分割等单任务识别方法为主。本文基于深度学习实例分割网络探索可同时完成滑坡目标定位和语义分割的多任务识别方法。首先，基于谷歌地球影像构建了包含3822个样本的黄土滑坡样本数据集；然后，采用单阶段实例分割网络(YOLACT)构建了基于小样本学习的黄土滑坡多任务自动识别模型；最后，通过大、中、小3种比例尺度的滑坡测试样本对识别结果进行评价。试验结果表明:①滑坡目标定位框(Box)平均精确度为61.66%，滑坡语义分割掩码(Mask)平均精确度为62.0%，大比例尺测试结果中Mask交并比为0.88；②基于YOLACT构建的滑坡识别模型可同时完成滑坡目标定位和滑坡高精度掩码分割的双任务识别，为滑坡多任务自动识别及快速制图提供了技术支撑。

Multi-task automatic identification of loess landslide based on one-stage instance segmentation network

Automatic landslide identification can solve the problem of slow speed of manual visual interpretation. The existing automatic identification methods based on deep learning are mainly single-task recognition methods such as object detection and semantic segmentation.In this paper, the instance segmentation network based on deep learning is used to explore a multi-task identification method that can achieve landslide target location and semantic segmentation simultaneously.Firstly, a dataset of 3822 loess landslide samples is constructed based on Google Earth images. Then,the multi-task automatic identification model of loess landslide based on small sample learning is constructed by using the one-stage instance segmentation network YOLACT. Finally, the identification results are evaluated by the large, medium and small scale landslide test samples. The results show as follows:①The average precision of landslide target positioning Box is 61.66%, the average precision of landslide semantic segmentation Mask is 62.0%, and the intersection over union of Mask in large scale test is 0.88. ②The landslide identification model built based on YOLACT can complete the dual-task identification of landslide target positioning and high-precision mask segmentation at the same time, which proride technical support for the automatic multi-task identification and rapid mapping of landslide.