Multi-resolution Remote Sensing Image Fusion Classification Based On Dual-Branch Neural Network

In the process of using remote sensing technology to detect the surface of the earth,the fusion classification of multi-resolution remote sensing images has always been a very critical research topic.On the one hand,with the development of information technology and hardware With the support provided,people can obtain higher and higher resolutions and content from various remote sensing platforms such as satellites and airplanes.More and more complex remote sensing images.The unique data of these remote sensing images is multi-source and heterogeneous,the target structure is changeable,and the background The complex sceneries and other characteristics make it more and more difficult for traditional methods to meet the demand for efficient interpretation.On the other hand,the development of deep learning in the fields of natural images,video and voice is in full swing,showing its Powerful feature extraction capabilities for massive data.Therefore,in the work of this article,we aim at The special nature of remote sensing data,a deep neural network model specially used to process remote sensing data is designed to complete Into the remote sensing image fusion classification task.The specific content includes:1.Propose a deep dense classification network based on adaptive sample neighborhood migration for high-resolution remote sensing image fusion classification task.The network first adaptively migrates the neighborhood of the center pixel to determine the range based on the characteristic information of the pixel to be classified,so as to obtain more sample blocks containing homogeneous information.Secondly,we introduce the concept of dense path,and effectively increase the transmission of gradient information in the network through dense connections on the basis of the two-branch network,and use the self-attention mechanism to enhance the feature representation in the process of extracting the deep features of the network.2.Aiming at the task of fusion classification of multi-spectral and panchromatic images,a space-spectrum collaborative fusion classification network is proposed.it It aims to integrate feature-level fusion and classification into an end-to-end network model framework.Consider the task is a big For remote sensing scenes of different sizes,we use the adaptive sample selection strategy proposed in the previous chapter.In the network structure,based on the image blocks captured by the sampling strategy,Spectral Data has designed a global channel attention module,which highlights the advantage of rich spectrum information of multi-spectral data;and A context space attention module is designed for panchromatic data,which highlights the advantages of high spatial resolution of panchromatic data.Then increase the information transmission path between the two branches to effectively reduce the difference in the characteristics of the branches,so that the two characteristics are merged with each other.Finally,the deeper features are extracted from the fused features for classification.The experimental results on the high-resolution remote sensing data set prove the effectiveness and robustness of the method.3.A Transformer model for multi-resolution fusion classification of remote sensing images is designed.The Transformer model can extract the features of the network more concisely by enhancing the effective representation of feature channels while maintaining the spatial structure characteristics of each sample block.In addition,the Transformer model breaks the tradition of using feedforward network structures such as convolution and pooling,and trains the network by decomposing and re-encoding the tiles through the Transformer model.Experiments have proved that this algorithm performs well in remote sensing image fusion classification tasks in large scenes.