Deep Learning Classification Based On Multi-source High-resolution Remote Sensing Images

With the rapid development of earth observation technology,sub-meter and even centimeter remote sensing images have gradually become popular.As an important method of observing human activities,high-resolution remote sensing images play an important role in land change monitoring,urban space planning,and military reconnaissance.The rich spatial detail information and weaker spectral information of high-resolution remote sensing images have brought great challenges to the efficient processing of remote sensing image semantic segmentation tasks.In the traditional remote sensing image semantic segmentation methods,the most commonly used is the pixel-based ground object information extraction method and the object-oriented information extraction method.Pixel-based information extraction method has been widely used in hyperspectral image classification,but the utilization rate of rich spatial information is almost zero,and it is not suitable for high resolution remote sensing images.Object-oriented information extraction methods rely more on expert knowledge by manually selecting feature parameters and image segmentation thresholds,and use shallow structures in model classification,which leads to poor classification effect.In recent years,Convolutional Neural Network has gradually become an important means for semantic segmentation of high-resolution remote sensing images by virtue of its high classification accuracy and automatic learning of data features from massive data without the need of experts or prior knowledge.However,the convolutional neural network also has some problems in semantic segmentation,such as rough classification results,obvious stitching marks and insufficient learning ability of multi-scale features.How to further improve the model performance still needs further study.In view of this,on the basis of previous studies,the paper fully excavates the feature information of high-resolution remote sensing images,combines multi-scale feature learning and attention mechanism to improve the U-Net network model,and proposes a new network structure.Experiments are carried out in tasks such as urban land use classification and pear t ree classification,and strive to improve the accuracy and efficiency of model classificati on,and provide a certain reference for remote sensing image extraction and other work.The main research of this article is as follows:(1)With the theme of semantic segmentation of high-resolution remote sensing images,we systematically introduce the progress of domestic and international research on semantic segmentation of multi-source high-resolution remote sensing data as well as research trends,and illustrate the advantages and disadvantages of traditional image segmentation methods and deep learning semantic segmentation methods,laying a good foundation for the urban land use classification task and pear tree classification in Yuli County in the later paper.(2)According to the characteristics of high-resolution images and by collecting corresponding images and GPS data,this paper constructed the land cover classification data set of Urumqi City and the pear tree data set of Yuli County.(3)To solve the problem of insufficient multi-scale feature learning ability of convolutional neural networks,we constructed a semantic segmentation network model(Multi-U-Net)that integrates multi-scale feature information and attention mechanism.First,the network improves the convolutional layer in the U-Net network by constructing a universal and extensible residual module under multisensory field(RMMF),which uses a variety of sizes of convolution kernels to learn images global and local information.At the same time,residuals are used to fuse the feature information of the shallow and deep layers,which effectively expands the network of receptive field scope and abate the degradation ability of network.In addition,the attention mechanism module is introduced to optimize the feature selection method of the model.By re-weighting each channel,the unimportant channel information is weakened and the important channel features are strengthened.(4)In order to test the performance of the model,experiments were carried out in remote sensing image datasets with different spatial resolutions,and more than five current mainstream methods were selected from each dataset for comparative analysis.The experimental results show that Multi-U-Net can achieve excellent classification accuracy in the semantic segmentation task of multi-source remote sensing images.It provides some reference value for the subsequent extraction of high resolution image ground object information.