Building Recognition And Change Detection From High-resolution Remotely Sensed Imagery Using Deep Learning

Accurately and rapidly obtaining buildings’ spatial location and change information from high-resolution remote sensing images plays an essential role in many practical applications,such as urban planning management,military detection,and disaster emergency response.However,it is still challenging for existing approaches to realize end-to-end recognition and change detection,mainly attributed to 1)building objects usually vary in type and scale and 2)the limited capacity of non-linearity representational learning.As a result,the contradiction between poor intelligence of interpretation methods and the ex-panding applications is becoming increasingly conspicuous.Recently,with the emerging of a new round of scientific and technological revolution and industrial transformation,the artificial intelligence technology represented by deep learning has brought a new reform in photogrammetry and remote sensing field.Inspired by these,to address the remaining shortcomings of existing research,this dissertation conducted an in-depth study on ap-plying deep learning in building recognition and change detection from high--resolution remote sensing images,respectively.The main contents of this study are as follows:1.Considering the building recognition in shaded remote sensing images with high resolution,based on the non-local spatial shadow context,a novel shadow detection method is developed to accurately locate the shaded regions for recovering the degraded building information caused by shadow.First,a non-local context aggregation method is proposed to effectively capture beneficial shadow semantic associations.Then,it is embedded into the modified U-Net network to produce accurate shadow maps.Compared with the pre-vious methods,the proposed model has tackled shadow semantic confusion well.In ad-dition,an extensive remote sensing shadow data set is constructed to support the corre-sponding research.Experiments demonstrate that the proposed method can boost building extraction when copping with the shaded images with high-resolution.2.A novel scale- and edge-aware building extraction method is proposed.Under the promotion of denser spatial pyramid pooling,refined multi-level feature fusion,and multi-task driven refinement,the proposed method is cable to produce satisfactory build-ing recognition results.First,the Denser Spatial Pyramid Pooling module,derived from the Atrous Spatial Pyramid Pooling,is developed to enhance the ability to capture the multi-scale features of building objects.Next,to make better full use of the low-level fea-tures with high-resolution,an Attention-guiding Refinement Unit is designed to refine the low-level features before further fusion.Finally,multi-edge-aware tasks are embedded to reinforce the network to pay more attention to the hard examples of building boundaries during training.Compared with the previous methods,the proposed method has filled the gap that the existing deep-learning-based building recognition methods can not obtain sat-isfactory results.Meanwhile,the performance outperforms other state-of-the-art semantic segmentation and building recognition methods.3.A real-time lightweight building recognition method is proposed to address the shortcoming that the inference speed of the existing deep-learning-based methods is slow.First,the conventional Basic Residual Block is modified by introducing a multi-path struc-ture? Then,a lightweight backbone network is developed based on the modified residual block.Next,following the ”training to deployment” concept,an identity transformation is constructed to decouple the model during training and inferring.Finally,the siamese net-work is introduced to build the bilateral backbone.Compared with the traditional method,the proposed method makes full use of the training advantage of the multi-path network and benefits from the fast inferring speed of the single-path network.Experiments show that the proposed method has maintained the best trade-off between accuracy and inferring speed and can meet the demand of real-time building recognition.4.Considering the two issues: 1)the amount of commonly used data sets for build-ing change detection is insufficient,and 2)the class imbalance is prominent,a high-performance and lightweight building change detection model is developed with foreground-aware optimization.First,a lightweight backbone network is proposed to encode the in-depth features by grouping the residual block.Then,based on comparative learning,a lightweight difference learning architecture is designed.Next,a foreground-aware opti-mization method is introduced to handle the class imbalance problem in building change detection by dynamically re-weighting the training loss.Finally,some factors that might influence the accuracy of deep-learning-based building detection and the applicability of different methods are discussed.Experiments reveal that the accuracy and generality of the proposed method significantly surpass the other building change detection methods and the state-of-the-art change detection methods.