Remote Sensing Image Building Extraction Method Based On Deep Learning

With the advancement of remote sensing satellite technology,the earth observation systems based on high-resolution remote sensing images have increasingly become perfect.Automatic building extraction based on remote sensing images has a wide range of application requirements in urban planning and disaster monitoring,and has always been an important direction in the field of computer vision,and has attracted a lot of research.From initial human visual interpretation to extraction using machine learning methods to the introduction of deep learning methods,researchers continue to refine the technology of automatic buildings extraction.The current research hotspots of building extraction mainly focus on how to improve the extraction accuracy of buildings and reduce the cost of extraction.However,the existing research has many problems such as difficulty in establishing feature associations,unbalanced feature utilization,and high cost of labeling datasets,resulting in many challenges in building footprint extraction.In view of the shortcomings of existing building extraction methods in terms of accuracy and time cost,this thesis proposes a series of deep learning models based on convolutional neural network,following three aspects:(1)Aiming at the problem that it is difficult for mainstream methods based on convolutional neural network to establish category feature associations due to the high resolution and complex background of remote sensing images,this thesis proposes a Ushaped feature extraction network based on deepened space module.The U-shaped feature extraction network uses the slow up-down sampling mode and the structure of long skipping connections to obtain feature maps at different depths,which reduces the information gap between high-and low-dimensional feature maps.The deepened space module is introduced to activate the important features of channel and spatial dimensions,and help the network establish category feature connections.The experimental results show that the proposed network structure and the deepened space module can reduce the error and omission extractions,and improve the extraction accuracy.(2)Aiming at the problem that the existing methods based on convolutional neural network cannot balance the use of features of buildings with large-scale differences due to the limitation of receptive fields,resulting in incomplete extraction results of large buildings,this thesis proposes an adaptive screening feature network based on encoderdecoder structure.Firstly,the up-sampling structure is optimized based on the deepened space module,and its versatility is extended to help the network locate buildings and construct boundaries.On this basis,this thesis proposes an adaptive information utilization block as the bottleneck block of the network to expand the receptive field,capture deep feature mappings,and introduce adaptive channel branch to independently screen effective information.The experimental results show that ASF-Net can completely extract large buildings while maintaining the detailed information of small buildings,so as to improve the overall extraction accuracy.(3)Due to the high resolution of remote sensing images,which leads to excessive cost of pixel-level labeling datasets,this thesis proposes a one-stage weakly supervised method based on class activation mapping.On the one hand,the classification branch calculates the category weights through the backbone network,then obtains the class activation maps,and generates pseudo-labels after refinement processing.On the other hand,the semantic branch introduces the bottleneck attention mechanism at the bottleneck of the backbone network to enhance the context information of the high-and low-dimensional feature maps.Then the joint loss function is introduced to cooperate with pixel-level cross-entropy loss to optimize the entire network.The experimental results show that the proposed weakly supervised network realizes the automatic building footprint extraction using image-level labels,and the accuracy is better than state-of-theart.