Object-based Image Classification Of Post-earthquake High Resolution Imagery Based On Deep Learning

The rapid and accurate recognition for targets objects after earthquake is crucial for the assessment of damage and the acquisition of damage information.In traditional method,fields' investigation is used to obtain information about the earthquake disaster.However,this traditional method consumes large amounts of manpower,material resources and time,which makes it unable to meet the requirements of emergency rescue of earthquake.Remote sensing techniques play an essential role in investigating damage information caused by earthquakes due to its prompt availability after disaster and wide coverage.The interpretation of remote sensing image is becoming a hotspot in the damage information extraction.This paper analyzes the current difficulties faced by the post-earthquake image segmentation first,and then introduces object-based post-earthquake high resolution image classification method using deep learning,and constructs an integrated process for quick and accurate extraction of post-earthquake disaster information.The main contents of this thesis are shown as follows:(1)Analyze the characteristics of typical targets in post-earthquake high-resolution images,characterize the damage objects,determine the spectral/spatial features suitable for the segmentation,and then apply the features to an adaptive spectral-spatial descriptor and combine it into a dynamic merge strategy to realize adaptive dynamic region merging.Especially,the new descriptor offers the adaptive semantic descriptions for geo-objects and thus is capable of characterizing different damage objects.Besides,in the dynamic region merging strategy,the adaptive spectral-spatial descriptor is combined with graph models to construct a dynamic merging strategy.The new strategy can find the global optimal merging order and ensure that the most similar regions are merged at first.With combination of the two strategies,ADRM can identify the spatially scattered objects and alleviates the phenomenon of over-segmentation and under-segmentation of post-earthquake high-resolution images,especially these characterized by large-scale,spatially distributed damages such as landslides and collapsed building areas.(2)Exploit the damage sample selection method and construct the damage training sample database.Instead of the pixel,the homogeneous regions in the multiscale segmentations obtained by object-based image segmetation methods are taken as the basic unit for the training samples selection and the training sample database construction.The results show that the method of selecting training samples based on the multi-scale segmentation results effectively solves the problem that the samples of complex damage targets are difficult to identity,and forms a new paradigm for the training samples selection in the post-earthquake high-resolution images classification.(3)Construct a convolutional neural network classification model to study the deep semantic features of typical objects such as landslides,intact houses,bare land,and so on.Quantitative and qualitative experiments show the great potential of CNN in post-earthquake high-resolution image classification,and also highlight the impact of CNN's reduced feature resolution on classification results.(4)A deep learning network that combines convolutional neural network and objectoriented segmentation is established to classify post-earthquake high-resolution remote sensing images,extract damage information,identify typical earthquake damage targets,and achieve rapid damage mapping.In this paper,the CMSCNN strategy is proposed in combination with CNN and ADRM,and realized the whole process of damage information extraction.Particularly,the CNN classification achieves the accurate classification of targets,while the ADRM multi-scale segmentation provides the precise boundaries of the geo-objects.The CMSCNN algorithm significantly improves the accuracy of the localization and classification,solving the limitation that the traditional CNN algorithm generally blurs the boundary in the high-resolution images classification.Quantitative and qualitative assessments demonstrate that CMSCNN is simple,practical,and suitable for rapid damage mapping.