Research On Building Detection From Remote Sensing Images With Deep Learning And Structural Priors

Buildings are the most common and changeable man-made land-covers in cities.The achievement of fast and accurate building detection has great significance in many practical fields,e.g.,urban planning,supervision of illegal buildings,and ecological environment protection.With the rapid development of remote sensing technology,the spatial resolution of the remote sensing image has been improved,meanwhile,earth observation has been developed from single-temporal static observation to multitemporal dynamic observation,introducing a new challenge for building detection.On one hand,the variable size of buildings and the complex surrounding environment make it difficult for accurate building detection.On the other hand,the multitemporal remote sensing images are always acquired at different times and imaging environments,further increasing the difficulty of changed building detection.In recent years,deep learning technology has been widely used in tasks of object extraction and recognition in remote sensing images and achieved promising performance,which is mainly due to its powerful capability in adaptively learning intrinsic features from images.However,the existing deep learning methods mainly rely on data-driven feature learning and ignore the structural prior knowledge of buildings,which still has difficulty in obtaining high-precision building detection results with clear edges.Therefore,this thesis attempts to integrate deep learning and structural prior knowledge to construct structural priors guided deep neural networks in building detection and changed building detection tasks,respectively.The details are as follows:(1)A building detection method based on convolutional neural network and structural prior is proposed to solve the problem that the edges of buildings in high-resolution remote sensing images are difficult to accurately extract.This method proposes an edge enhancement module to obtain more discriminative features.On the one hand,it can learn the offset field and constrain the flow of deep semantic information through shallow spatial features to achieve deep and shallow feature alignment.On the other hand,structural information can be obtained to guide building edge refinement.Moreover,the multi-level prediction probability maps are adaptively fused to further improve the detection accuracy.Experimental results on two building detection datasets show that the proposed method achieves better results in terms of evaluation metrics and visual comparison results.(2)Focusing on the problem that it is difficult to accurately identify changed buildings from multitemporal remote sensing images,a building change detection method based on recurrent convolutional neural network and structural prior is proposed.A difference analysis module is introduced to further produce discriminative features,which is constructed based on the basic long short-term memory module.In addition,both the discriminative features and the estimated edge structure information are jointly exploited to predict building change maps.The prior edge information is used to push the predicted changed buildings to preserve the original structure,which can further improve the accuracy of building change detection.Experimental results on two building change detection datasets demonstrate that the performance of the proposed method outperforms several state-of-the-art approaches,in terms of objective metrics and visual comparison results.(3)Based on the proposed remote sensing image building detection method,we analyze the courtyard closure of traditional Chinese courtyard residences in southern Hebei.Preliminary research shows that the traditional Chinese courtyard residences in southern Hebei are closing the original open courtyards with roofs of various materials.Due to the vast territory,manual exploration is extremely difficult,so we combine remote sensing observation with deep learning techniques to analyze the distribution of this phenomenon.This thesis designs a complete technical solution,including remote sensing image acquisition and preprocessing,training sample labeling and data augmentation,deep neural network training,prediction of traditional Chinese courtyard residences with different courtyard closure forms,model verification,and adjustment.Finally,the adjusted model is used for courtyard detection on large-scale data.According to the statistical results,the current characteristic of the distribution and future expansion trend of this phenomenon are analyzed.