Research On Deep Learning-based Hyperspectral Image Classification Methods And Their Applications

In recent years,hyperspectral image classification technology has become one of the important research areas in remote sensing image processing.Hyperspectral image data contains high-resolution spectral and spatial multidimensional information,which can provide rich spectral information of ground objects.Therefore,it has wide applications in fields such as land cover classification,resource and environmental monitoring,and agricultural ecology.However,processing and classification of hyperspectral image data face many challenges,such as insufficient feature extraction for fine details,significant spectral information loss,inadequate utilization of spatial information,and failure to combine spectral and spatial information effectively.To address the shortcomings of existing networks,this paper proposes two different hyperspectral image classification networks based on convolutional neural network(CNN)and fully convolutional neural network(FCN)as the underlying network architecture.Various comparative and ablation experiments were conducted on multiple public datasets and a real honey pomelo orchard dataset.The results show that the hyperspectral image classification networks proposed in this paper have certain advantages compared to other currently popular networks.The main contents of this paper are as follows:(1)In response to the problem that the current research on hyperspectral image classification has not fully considered the difference in feature extraction corresponding to the size of ground objects in joint spectral-spatial feature extraction,we propose a new multi-scale spectral-spatial attention network(MS~3A-Net)based on a CNN network.The front-end of the network uses a spectral-spatial cooperative attention(SSCA)block,and the feature extraction part uses a multi-scale attention feature extraction(MAFE)block in the pyramid bottleneck residual structure to perform spatial-spectral feature extraction.This can extract features of ground objects of different scales and minimize interference from ground objects of different scales.MS~3A-Net can enhance the influence of effective pixels in spatial and spectral dimensions,suppress the influence of invalid pixels and even interference pixels.Through experiments,it is demonstrated that the proposed method has certain advantages in the classification performance of the network compared to the SVM method and other CNN-based hyperspectral classification methods.(2)In response to the difficulty of obtaining global information from images using CNN-based local learning methods and the challenge of adapting to various input-output image sizes due to highly overlapping patches,this paper proposes a global learning network based on large receptive fields(GLN-LRF)structure,which adopts an encoder-decoder model with skip connections as the basic framework.In the downsampling part of the encoder,multi-scale contextual feature information is extracted through the fusion of large convolution kernels,spatial separable convolution,and dilated convolution in the large receptive fields context exploration(LRFC)block.In order to extract deep semantic information,the paper proposes a multi-scale simple attention(MSA)block in the decoder to achieve deep feature fusion.Experimental results on various high-spectral datasets with different spatial scales demonstrate that the improved classification method has significant advantages compared to current CNN-based and FCN-based classification networks.(3)To address the high spectral image classification needs in real agricultural scenarios,we collected data from Pinghe honey pomelo orchards using airborne hyperspectral imaging equipment and created a corresponding dataset.In this dataset,we conducted a detailed comparative analysis of our proposed MS~3A-Net and GLN-LRF methods with other current classification methods.