Research On Convolutional Model-Based Hyperspectral Image Classification Algorithm

With the continuous development of remote sensing technology,hyperspectral image classification has become an important research direction in remote sensing applications.Hyperspectral images contain a large amount of spectral and spatial information,with high resolution and spectrum integration advantages,and have received wide attention.However,the high dimensionality of hyperspectral data,strong correlation between bands,and limited training samples pose certain challenges to the extraction of spectral spatial features.In recent years,convolutional neural networks(CNNs)have been widely used in hyperspectral image classification tasks due to their powerful learning ability,which can better extract nonlinear and deep-level features of hyperspectral images.However,in addition to the difficulties inherent in hyperspectral images,hyperspectral images with different spatial resolutions also have different problems.Therefore,targeted algorithms need to be selected in the classification process to improve classification accuracy and computation efficiency.This article focuses on how to fully utilize the spatial structure information and spectral structure information of hyperspectral images to efficiently improve the accuracy of land classification and covers the following aspects:(1)For small hyperspectral datasets with low spatial resolution and few labeled samples,existing convolutional deep learning models have the problem of large parameter quantities,long training times,and insufficient utilization of spectral and spatial information,resulting in limited performance.This article proposes a lightweight hybrid convolutional network based on dense connection structure and attention mechanism.The network first uses a 1×1 convolution kernel to reduce the spectral dimension and uses a dense connection network structure to fuse the spectral information in the spectral information extraction part.It also introduces dilated convolution to expand the spectral dimension to extract more comprehensive spectral features.Secondly,to strengthen joint feature extraction of spatial and spectral features and reduce computational costs,separable convolution is applied.Finally,to reduce the impact of noise on classification,a spatial-spectral attention module is used,which can enhance the weight of effective features and suppress the weight of ineffective features,making the classification result more accurate.The network uses 2D-CNN to implement classification,greatly reducing network complexity and improving network performance,achieving better classification results.Experimental results were obtained on three small datasets,namely Inpines Pines,Pavia University,and Salinas Valley.The OA achieved were 99.24%,99.92%,and 99.96%,respectively,while the average AA were 99.15%,99.89%,and 99.96%.The Kappa coefficients were 0.9913,0.9990,and 0.9996 for the three datasets,respectively.(2)For hyperspectral images with diverse land cover types,rich pixel information,and high spatial resolution,the model’s computational complexity for processing and extracting land features is increased.At the same time,the expression of spatial texture and geometric information in hyperspectral images is weak,and a single convolutional scale is difficult to fuse contextual information.This article proposes a multiscale and multiresolution attention feature fusion convolutional network based on wavelet transform.The network uses wavelet transform for four-level decomposition on the spectral band to extract spectral features layer by layer and reduce computational complexity.Meanwhile,a spatial information extraction module is designed,which uses a pyramid attention mechanism and utilizes multiscale features through inverse jump connection network structure to enhance the expression of spatial texture,thus improving classification accuracy.This network effectively solves the problems of single-scale feature extraction and ignoring spatial texture details in traditional 2DCNN feature extraction,improves feature fusion effect,and performs well in hyperspectral image classification tasks.Experimental results were obtained on two high spatial resolution hyperspectral datasets,namely WHU-Hi-Han Chuan and WHUHi-Hong Hu.The OA achieved were 99.64% and 99.67%,respectively,while the AA were 99.72% and 99.73%,respectively.The Kappa coefficients were 0.9934 and0.9932 for the two datasets,respectively.