Remote Sensing Object Identification And Rapid Seismic Landslide Assessment Based On Deep Transfer Learning

Large-scale earthquakes can trigger a series of hazards,especially the seismic landslides,causing tremendous damages to urban and rural areas.“Where the landslide occurred after the earthquake” and “How about the scale and hazard of the seismic landslide” are the most concerned about the seismic landslide disaster.The occurrence of seismic landslides is the first step in the emergency response of earthquake.With the development of remote sensing technology,in order to to solve the problem of where landslides are after the earthquake,the automatic recognition and segmentation of seismic landslides based on the spatial perception of remote sensing images has become the basic requirement.This used to be conducted through pixel-based or object-oriented methods.However,these methods fail to develop an accurate,rapid,comprehensive,and cross-scene solution for seismic landslide recognition due to the massive amount of remote sensing data,variations in different earthquake scenarios,and the time sensitivity for post-earthquake rescue.On the other hand,under the background of remote sensing data with large quantities,the core problem of the scale and hazard of seismic landslides is how to integrate remote sensing and geographic information spatial cognition,that is,how to use the prior data and model of regional landslide assessment to provide rapid and accurate landslide disaster information.In the past,the general methods of image processing are simply applied to remote sensing images of disaster.The spatial-temporal particularity of remote sensing data and other disaster geographic information attributes are not fully coupled.Therefore,the efficiency and accuracy of remote sensing cognition have fundamentally limited the application of remote sensing data in the seismic landslide rapid identification and assessment.In addition,with the rise of artificial intelligence and machine learning,where the seismic landslide(remote sensing image spatial perception)and how the seismic landslide(geographic information spatial cognition)are explored in this thesis.Further,after the earthquake,within 10 minutes or even 5 minutes,the accurate landslide range,landslide risk,and loss are accurately and robustly evaluated after the earthquake.The primary contents of our proposed framework are summarized as follows:(1)In order to solve the spatial accuracy and cross-scene robustness of regional landslide identification,including supervised and unsupervised remote sensing object recognition,a comprehensive and general scheme for recognition of remote sensing objects is proposed based on convolutional neural network and unsupervised domain adaptive transfer learning.Specifically,When the corresponding labeled landslide training samples are available in the study area,a deep supervised segmentation model,dubbed MFFENet is proposed to extract and fuse the multi-scale features of objects in remote sensing images.The proposed MFFENet achieves a competitive segmentation accuracy on the public semantic benchmark data.However,in most cases,the suddenness of earthquakes and the massive amount of data in remote sensing images make the labeling-all task difficult.Based on adversarial learning method,the Adversarial Domain Adaptation Network(ADANet)in the output space domain adaptation and the Class-aware Generative Adversarial Network in the feature space domain adaptation are proposed to enhance the adaptability of the segmentation model.Experimental results demonstrate the effectiveness of ADANet and Ca GAN.Furthermore,the multi-level output space adaptation scheme proposed in ADANet and the class-aware domain alignment scheme in Ca GAN are proved to further improve the adaptability of the segmentation model.Finally,we preliminarily discuss the situation that the annotated source domain cannot be found due to the lack of source domain data or the mismatch of category,and a domain adaptation model without source domain is proposed;(2)In order to further improve the accuracy of the results of supervised or unsupervised segmentation models,a fine-tuning module is proposed,which includes Geologic Feature Fusion(GFF)and Bi-temporal Changing Feature Fusion(Bi CFF).The GFF module is proposed to eliminate these non-landslide areas(such as roads,small collapse,etc.)that can be distinguished by landslide geological features.The Bi CFF module is proposed to eliminate the irrelevant areas of the earthquake,such as historical landslides and some unrelated targets;(3)In order to solve the rapid integration of remote sensing and geographic information system,the spatial geographic information attribute of seismic landslide is conducted based on geographic information system.In addition,with the help of other non-remote sensing data and landslide risk assessment prior models,each landslide static attribute in the region is extracted to reflect the landslide distribution,risk situation and loss situation.Finally,a comprehensive seismic landslide identification and assessment scheme is proposed from the perspective of spatial granulation and attribute granulation;(4)The proposed scheme is applied in two earthquake-triggered landslides in Jiuzhaigou(China)and Hokkaido(Japan),using available pre-and post-earthquake remote sensing images.These experiments show that the supervised framework(MFFENet model with fine tuning module)has the most advanced performance in regional landslide identification(the comprehensive accuracy is 91.59 %).Unsupervised framework(ADANet model with fine tuning module)shows excellent and robust performance in different seismic landslide tasks(overall accuracy 82.33 %).Furthermore,in order to realize the rapid assessment of regional earthquake landslides,the spatial granulation of the supervised or unsupervised framework recognition results is conducted.The spatial geographic information system is integrated with the consideration of landslide scale and landslide hazard degree.Overall,through the geographic information space system,the rapid assessment framework of seismic landslide based on deep transfer learning can realize the spatial granulation of remote sensing and the efficient integration of landslide risk attribute granulation.As a result,the attribute feature space and regional landslide database of seismic landslide are established.Finally,a feasible way for intelligent emergency relief of earthquake landslides is explored on the basis of fully understanding the particularity of remote sensing interpretation,visual cognition and geological analysis technology.