Deep Learning Based Hyperspectral Image Classification Algorithms With Small Samples

Hyperspectral image classification is a crucial component of remote sensing technology,with significant value in various applications,including precision agriculture,geological exploration,and medical diagnosis.In practical applications,the lack of large,publicly available labeled datasets for hyperspectral images may result in a shortage of training samples.In addition,the hundreds of near-continuous spectral bands of hyperspectral images produce a large amount of redundant information while providing a rich set of features for classification.Therefore,the current challenge in hyperspectral image classification is how to extract valuable information to improve accuracy with small samples.Traditional machine learning-based classification methods are generally based on manual features and shallow learning,which are not very effective in classification.In contrast,deep learning-based methods can automatically learn deeper feature expressions and effectively improve classification accuracy.Based on the above analysis,this thesis conducts a study on the following three aspects based on a multi-branch deep learning approach.(1)The D3 DTBTA network structure is proposed by combining deformable 3D convolution and the three-branch triple attention mechanism.The network extracts features independently in three branches: the spectral branch,the spatial-X branch,and the spatial-Y branch,and applies an attention mechanism on each of the three branches in their directions to refine the features.The features extracted on the three branches are then fused for classification using a concatenation operation.In the network structure of D3 DTBTA,deformable 3D convolution is used to enhance the deformation modeling capability of 3D-CNN.Experimental results show that the D3 DTBTA network effectively improves the classification accuracy over the comparison methods in the case of small samples.(2)The D2Net＿TBTA network structure is proposed by combining Dense2 Net and the three-branch triple-attention mechanism.The network employs the Dense2 Net bottleneck module on the spectral branch,spatial-X branch,and spatial-Y branch for its design and implementation.Dense2 Net is a novel deep learning architecture that leverages the strengths of Res2 Net and Dense Net.It effectively extracts features at multiple scales while avoiding gradient disappearance and accelerating feature propagation.Then,the features extracted from the three branches are refined by the triple attention mechanism and fused for classification.Experimental results show that the D2Net＿TBTA network is effective in improving the classification accuracy with small samples compared with several other deep learning-based methods.(3)The Conv LSTM＿TBTA network structure is proposed by combining the convolutional long and short-term memory network(Conv LSTM)and the three-branch triple-attention mechanism.The network applies three Conv LSTM network layers and spectral attention block on the spectral branch to extract spectral feature.On the spatial-X and spatial-Y branches,a 3D-CNN-based Dense Net module and an attention mechanism are applied to extract spatial-X feature and spatial-Y feature respectively.Subsequently,the features extracted on the three branches are fused for classification.The experimental results show that the Conv LSTM＿TBTA network effectively improves the classification accuracy with small samples compared with the comparison methods.The experimental results demonstrate that the three network models proposed in this thesis have enhanced the classification accuracy even with limited samples,when compared to other existing deep learning methods.This provides a novel approach for the research of hyperspectral image classification.