Research On Building Shadow Extraction Method Of Remote Sensing Image Based On Deep Learning

In recent years,the spatial resolution of remote sensing images has been continuously improved,and the amount of data has increased dramatically.The rich feature information has provided massive data for urban planning,remote sensing mapping,disaster monitoring and other fields.Building shadows are common in high-resolution remote sensing images of cities.The presence of shadows will cause a large amount of information to be lost in the local area of the image,and the image quality will be reduced,which will affect the accuracy of further applications in the future.In addition,shadows are important and useful information for remote sensing images.Shadows can be used to retrieve building heights,which is beneficial to obtain building location information and geometric information.The traditional shadow extraction method is mainly based on shadow spectrum,geometry,texture and other features.This method requires a large amount of prior knowledge from the outside world,and too much manual intervention,resulting in low accuracy and efficiency of shadow extraction in massive data,Can not better meet the needs of actual engineering applications.Therefore,the accurate and efficient automatic extraction of building shadows from remote sensing images has important research significance and application value.This paper use deep learning technology to perform pixel-level semantic segmentation of building shadows.In view of the characteristics of large amount of data and complex features of remote sensing images,based on the encoder-decoder architecture of the classic image semantic segmentation model U-Net,this paper proposes an attention-based building shadow segmentation network Building Shadow Segmentation Network,ABSS-Net),and experiment by establishing shadow data sets of remote sensing images to verify and evaluate the effectiveness of the network.The main contents of this article are as follows:(1)The experimental data includes two types of data: remote sensing satellite QuickBird image and aerial image.The data set is produced by manual annotation,and the aerial image data set is divided into three different levels of 2000,6000 and 10000,which is convenient for later comparative experiments.In order to improve the generalization ability of model training,data preprocessing processes such as Gaussian filtering,histogram equalization and data enhancement are designed.(2)A feature extraction module based on the attention mechanism is added to the downsampling of the network encoder to strengthen the model's full expression and understanding of the extracted features.In addition,in the downsampling and upsampling feature fusion stage,a feature map fusion algorithm based on attention mechanism is designed to enhance the data flow of shallow and deep features.(3)Aiming at the imbalance between positive and negative samples in the data set,an improvement is made on the basis of the cross-entropy loss function of binary classification,and a balance factor is added to suppress the excessive performance of the model on negative samples.When predicting large images of any size,to prevent inaccurate edge prediction,a dilation prediction method based on boundary filling is proposed.(4)In order to verify the effectiveness of the method in this paper,the qualitative and quantitative accuracy evaluation of the method and U-Net model on remote sensing satellite images and aerial images are carried out.In addition,in order to verify the effect of this method on different magnitude data sets and different data sources,corresponding comparative experiments and accuracy evaluations were carried out.The experimental results show that the ABSS-Net classification result proposed in this paper is better than the classic model U-Net,and the average accuracy index is improved by about 4 percentage points,and the overall accuracy is 98%.In addition,the ABSS-Net model can effectively detect the target area where shadows are projected on buildings and vegetation,and can better distinguish between building shadows and water bodies.The research results prove the effectiveness of the method in this paper,and it has certain practical value in scene application.