基于Faster R-CNN目标检测的滑坡隐患识别——以福贡县城区为例

Landslides risk identification based on Faster R-CNN-a case study in Fugong county

滑坡是我国地质灾害中占比最高的灾害类型,由此造成的财产损失和人员伤亡巨大,快速全面识别滑坡隐患区域对防灾减灾工作意义重大。基于高分一号影像,采用Faster R-CNN算法构建了滑坡隐患自动识别模型,利用AP(Average Precision)值对其准确度进行了评价,最终测试结果显示AP值达到了92.42%;以福贡县城区为研究对象,分别采用基于Faster R-CNN和面向对象的滑坡隐患识别方法对该区域进行了滑坡隐患识别,结果表明:前者共识别出隐患点193处,正确解译176个,正确率为91.19%;后者共识别出245个滑坡隐患点,正确解译201个,正确率为82.04%;可见,基于Faster R-CNN的滑坡隐患识别方法的准确度高,其对地处高位、交通不便、隐蔽性强的滑坡隐患排查具有重要的指导意义。该方法虽对地表特征明显的滑坡具有较高的识别率,但对形态不完整、变形迹象不明显的滑坡难以识别,因此未来可考虑将其与InSAR、LiDAR技术相结合,通过综合研判进一步提高滑坡隐患识别的准确度。

Landslide is the dominated type of geological disaster in China,resulting in huge property losses and casualties.The rapid and complete identification of the landslide risk is of great significance to perform the disaster prevention and mitigation work.Based on high definition of No.1 image,the Faster R-CNN algorithm was used to build the automatic identification model for the landslide risk identification,and its accuracy was evaluated by using the AP(Average Precision)value.The final test results show that the AP value reached 92.42%.A case study was carried out in Fugong county,the Faster R-CNN and object-oriented landslide hazard identification methods were adopted respectively.The results show that with the former,193 risk points were identified,among which 176 points were correctly interpreted,resulted in an accuracy of 91.19%.With the latter,245 landslide risk points were identified,among which 201 points were correctly interpreted,resulted in an accuracy of 82.04%.It can be seen that the landslide risk identification based on Faster R-CNN is of higher accuracy,which has important guiding significance for checking and investigating landslide risks within the areas having high altitude,poor traffic conditions and high concealment.Although this method presents a high performance for identification of landslides in the area with obvious ground characteristic,it is difficult to identify landslides with incomplete morphology and little deformation marks.Therefore,in the future,the technology combining InSAR with LiDAR should be taken into account for further improvement on the accuracy of landslide risk identification via the comprehensive analysis.