Research On Hyperspectral Image High-Precision Classification Based On A Small Number Of Samples

In recent years,hyperspectral image classification methods based on deep learning models have been extensively studied because of their superior performance.Some excellent hyperspectral image classification methods have achieved very high classification accuracy by relying on enough training samples.However,in the case of a small number of training samples,there are still many problems such as insufficient feature extraction’s capabilities,serious overfitting problems,and unstable effects that lead to poor classification results.In the deep learning models,the hyperspectral image classification methods based on the combination of the attention mechanism and the convolutional neural network can often achieve higher accuracy.However,the applied spatial attention mechanism and channel attention mechanism are separated from each other,that is,the spatial attention model only extracts the correlation information within the spatial dimension,and the channel attention model only pays attention to the correlation between different channels.sex.Moreover,the applied attention mechanisms are also one-dimensional,two-dimensional or a combination of the two.In addition,the general convolution used for feature extraction is limited in the richness and hierarchic nature of features that can be extracted when there are fewer training samples.To alleviate these contradictions,this work mainly introduces two new hyperspectral image classification algorithms.The specific contents are as follows:(1)This paper proposes a new Dual-Triple Attention Network(DTAN),which uses a small number of training samples to achieve high-precision classification of hyperspectral images on the basis of capturing cross latitude’s interactive information.DTAN is divided into two branches to extract hyperspectral spectral information and spatial information respectively,which are called spectral branch and spatial branch.While applying the channel attention model to the spectral unit,the cross-dimensional interaction between the channel and the spatial domain is constructed.When the spatial attention model is used to the branch,the correlation with the channel domain is also considered.Moreover,DTAN introduces an efficient channel attention(ECA)module into the Dense Net,which allows the Dense Net to achieve partial cross-channel interaction.A series of experiments hanve proved that DTAN has a significant advantage compared to other models when the training samples are minimal.The effectiveness of feature extraction network and attention module ensures high-precision classification.(2)Although DTAN has a high classification accuracy,its time cost is slightly higher than that of the contrasting methods.This paper proposes a new Py-Cond Convolution Attention Network(PCAN),which further improves the efficiency of the model without decreasing or even increasing the overall accuracy.PCAN can fully and effectively capture the information of hyperspectral input data through convolution kernels with different weights and different scales.And for the first time,the attention of three-dimensional weights is applied to hyperspectral data,and satisfactory results are achieved on very small sample training sets.Specifically,PCAN is composed of spectral branch and spatial branch.The spectral branch uses Conditionally Parameterized Convolutions(Condconv)to customize the convolution kernel according to the input features,and directly generates three-dimensional weights to quantify the importance of the entire feature map.The spatial branching uses Pyramidal Convolution(Py Conv),which not only includes convolution kernels of different spatial sizes but also different depths,and can capture different levels of detail information in hyperspectral images.At the same time,competition and cooperation between channel information are captured through gating adaptation and channel normalization.A series of experiments showed that PCAN achieved a significantly better classification performance than the contrast models under a limited training samples,while maintaining a lower time cost.