Design And Implementation Of Landscape Classification Algorithm For Hyperspectral Images

Hyperspectral images contain both rich spectral and spatial information,which greatly enhances the ability to classify and identify features.With the development of deep learning technology,the research of hyperspectral image classification gradually brings in intelligent learning theory.In this regard,the application of neural networks has significantly enhanced the classification performance of hyperspectral images.Neural networks allow the extraction of more abstract features,but there are two intractable problems as follows: To begin with,although CNN-based models have achieved good results in HSI classification,the translation invariance and local connectivity of CNNs can affect the class effect of HSI classification.In Addition,deep learning models need sufficient labeled samples for parameter updates,but obtaining high-quality labeled samples is often time-consuming and laborious.It has become one of the research hotspots in the field of hyperspectral remote sensing in recent years to achieve high accuracy HSI classification results with a small number of marker samples.Hence,two solutions are proposed in this paper to address the above problems respectively,and the main work and innovations accomplished in the paper are as follows.Firstly,for the problems of the existing deep network model design workload,translation invariance and local connectivity of CNN,a hyperspectral data classification method based on multilayer perceptron combined with residual learning is designed.The method starts with the design of an improved Multi-Layer Perceptron model.The network training performance was enhanced by using deep hyperparametric convolution without increasing the inference computation.The Focal Loss loss function is utilized so that the network learns more useful HSI information.Then a non-linear and better h-swish activation function is introduced to improve the generalization performance of the network.At last,IMLP is inserted between two convolutional layers in the normal residual block to form the IMLP-Res Net model.This methodology can significantly improve the ability of the network to extract and generalize HSI features.The comparison experimental results confirm that the proposed method in this paper further boosts the classification accuracy of HSI.Secondly,to tackle the problem of low HSI classification accuracy under limited sample conditions,this paper designs a metric metric-based metric learning method for the classification of hyperspectral image features with joint spatial-spectral migration.In order to fully extract HSI fine features,a spatial-spectral feature migration network module is designed,and a SE-Net attention mechanism is introduced in the spectral dimension channel to selectively extract useful features and improve the sensitivity of the network to information features.Next,the spatial feature extraction part is implemented by the model parameters pre-trained on the HSRS-SC dataset to transfer the spatial feature knowledge to learn,and then extract the higher-order abstract features and thus mine the eigenattributes of the features.At the end,a gated feature fusion strategy is introduced,which can connect the extracted spectral-spatial HSI feature information to extract richer feature information.Experiments show that this method can effectively improve the feature classification accuracy of HSI under limited sample conditions and has good generalization ability to different target datasets.