Extraction Of Built-up Areas From High-resolution Satellite Imagery Using Deep Learning

Built-up areas are the main place where humans are engaged in production and life.The information on the location and distribution of built-up areas plays an important role in the real-time update of the basic geographic database,urban dynamic monitoring,urban planning and other applications.With the development of earth observation technology,the spatial resolution of remote sensing images has reached the sub-meter level,providing an effective data source for the acquisition of fine-scale built-up areas information.However,the improvement of the resolution also makes the features of the built-up areas in the image more complex.The traditional methods using handcrafted features have poor robustness and low generalization ability,and it is difficult to adapt to large-scale areas,especially complex scenes.In recent years,the rapid development of deep learning has provided a new paradigm for the recognition of built-up areas from high-resolution remote sensing image.Using deep learning can automatically learn features from sample data to achieve hierarchical representation of image features.Therefore,the method based on deep learning can better obtain the discriminative characteristics of the built-up areas/non built-up areas,achieve more accurate target detection,and have stronger robustness and generalization ability for large-scale complex scenes.Based on the related models and algorithms of deep learning,this paper deeply researches the automatic extraction method of built-up areas from high-resolution remote sensing images.The main work is as follows:(1)Facing the recognition of built-up areas from large-scale high-resolution images,this paper proposes a novel method for built-up areas extraction from high-resolution remote sensing images using convolutional neural network.In this method,the image blocks divided by grid are regarded as the basic unit of processing,and the convolution neural network based on dense connection and attention mechanism is used to classify the image blocks.Furthermore,the refined extraction of builtup areas is achieved by integrating the classification results of multi-scale grid translation.(2)By incorporating multi-scale dilated convolution,a full-convolution semantic segmentation algorithm is proposed to obtain more precise positioning of the built-up areas.In this method,the full convolution neural network is used as the network backbone,and the five convolution modules of FCN are used to extract features;then the multi-scale dilated convolution module is used to obtain more detailed context information;finally,the low-level features and high-level features are fused to obtain the saliency map of built-up areas with the same size as the original image,and the pixel-level semantic segmentation of built-up area is realized.(3)A dataset of built-up/non-built-up areas from high-resolution remote sensing image is constructed.Aiming at the task of built-up area extraction,two sample data sets of built-up areas at block level and pixel level are constructed by using the Gao Fen-2 satellite image of Shenzhen City,which are used to train the two different extraction models.(4)The performance of the proposed built-up areas extraction model is tested and evaluated on the high resolution satellite images of Shenzhen City,and good experimental results are obtained.Compared with the existing methods,the proposed method has better performance.Furthermore,using the proposed model,this paper realizes the fine detection of the whole area in Shenzhen,which shows that the model has the potential to achieve high-precision and high-resolution mapping of large-scale built-up areas.