Extraction Of Landslides From High Resolution Remote Sensing Images Based On Deep Learning Of Different Levels Of Supervision Information

As a common and frequent geological disaster in mountainous areas,landslides pose a serious threat to the construction of major projects and the safety of people’s lives and properties.Therefore,carrying out a comprehensive survey of landslide geological hazards in key mountainous areas to find out the main spatial distribution of landslides can provide a strong guarantee for local governments to establish and improve the geological disaster prevention,mitigation and relief system.It is difficult to obtain landslide distribution information in complex and difficult mountainous areas efficiently by traditional ground field survey methods,and it is difficult to obtain increasingly convenient high-resolution remote sensing images and image information processing and analysis technologies represented by Convolutional Neural Networks(CNN)for disaster monitoring.And high-precision information extraction provides a feasible technical means.The convolutional neural network model requires a large number of manually annotated samples for training.The scene-level and frame-level sample annotations are simple and easy to obtain,but the landslide extraction accuracy is low and the result granularity is too large for quantitative analysis,and pixel-level sample annotation is difficult to obtain.However,high-precision landslide information can be obtained.Focusing on the characteristics of landslides in high-resolution remote sensing images,this thesis studies high-precision pixel-level landslide information extraction methods under scene-level,frame-level and pixel-level sample supervision information.Specifically,the research content and results of landslide information extraction in this paper can be summarized as follows:(1)For scene-level samples,a weakly supervised landslide extraction method based on deep attention and multi-feature fusion is proposed.The method in this thesis solves the demand for pixel-level labeled samples of the semantic segmentation model by obtaining highquality class activation maps(CAMs)from the convolutional neural network model,and then trains a high-performance landslide extraction model.Focusing on the characteristics of landslides in high-resolution remote sensing images,this thesis proposes a landslide scene classification network based on deep attention and multi-feature fusion to improve the accuracy of CAM and the level of detail of landslide information.The fully conditional random field(F-CRF)algorithm optimizes the edge of the landslide disaster area to obtain higher-precision pixel-level pseudo-labels,and then uses these pseudo-labels to train the pixellevel landslide information extraction network in a fully supervised manner.The experimental results show that the landslide extraction accuracy of this method is significantly improved and the results are more refined,and the main accuracy indicators are better than the mainstream weak supervision methods,and achieved close to the strongly supervised model.(2)For frame-level samples,a weakly supervised landslide extraction method based on object detection and network transfer learning is proposed.The output result of the target detection network is landslide boundary information,which cannot provide landslide boundary information.In this thesis,we first explore strategies to obtain CAMs from several typical object detection models,and obtain high-quality pixel-level pseudo-labels by optimizing the generated CAMs.Then,on this basis,a weakly supervised remote sensing image landslide extraction framework based on network transfer learning is proposed,which effectively reuses the network parameters of the target detection model through transfer learning,and uses the obtained pixel-level pseudo-labels to fine-tune the semantic segmentation branch,so as to obtain pixel-level extraction results.The experimental results show that the method obtains accurate pixel-level landslide information,which is also better than the traditional weak supervision method under the supervision of bounding box information based on graph segmentation,and the main accuracy indicators are significantly improved,and achieved results very close to the strongly supervised model.(3)For pixel-level samples,a fully supervised landslide extraction method based on hard example mining and instance segmentation network is proposed.Taking pixel-level supervision information as an example,this thesis attempts to study the problem of false positives in landslide target recognition from two aspects: sample information expansion of difficult examples and optimization learning of difficult examples.First,focusing on the characteristics of landslides in high-resolution remote sensing images,a copy-paste-based landslide sample data expansion and enhancement method is proposed to enhance the richness and diversity of the sample data set.Then,by introducing the idea and method of hard case mining,a landslide extraction model based on hard case mining and instance segmentation network is designed to improve the contribution of difficult samples to gradients,thereby enhancing the recognition ability and robustness of the model to landslide targets.The experimental results show that both of them can effectively reduce the wrong prediction,and provide a scheme reference for the accurate identification and wide application of landslide targets in large-scale remote sensing images.