Hyperspectral Image Classification Based On Multi-scale Feature Transfer Learning And Deep Neural Networks

Hyperspectral images contain rich information and have the feature of high spectral resolution and high spatial resolution,which helps people constantly recognize the essential characteristics of ground object attributes.As an important application in the field of hyperspectral remote sensing,the classification technology of ground objects in hyperspectral images aims to accurately classify the classes of ground objects covered by hyperspectral images.Although the rich spectral and spatial information in hyperspectral images brings rich information,there are still problems such as band redundancy and "dimension disaster" caused by the large amount of information.At the same time,the phenomenon of the same spectral for different surface features and different spectra for the same surface feature in hyperspectral image classification also restrict the classification accuracy of hyperspectral images.As a common method of hyperspectral image classification,supervised classification m ethods'performance rely on a large number of labeled hyperspectral training samples.To solve these problems,this article introduce the multi-scale feature extraction and transfer learning based on the model into the classification of hyperspectral image.Systematically introduces the basic principle of multi-scale feature extraction and transfer learning,corresponding algorithm was designed and implemented.Four open hyperspectral dataset were used to verify algorithm performance.Experiment results show that the classification accuracy of hyperspectral images significantly improved.The main aspects of the study and results achieved are as follows:The scale of objects in hyperspectral images is quite different,lead to the problem of low classification accuracy of mixed pixels.In this article,a multi-scale feature of neural network based in feature pyramid is proposed.Through the "bottom-up" process,the input hyperspectral data is down-sampled,and the feature map containing deep abstract semantic information is output.The feature map output in the last step is up-sampled to obtain a feature map with high spatial resolution containing low-level semantic information which called "top-down".The feature maps of these two processes are combined to obtain a feature map that contains both low-level semantic information and high-level semantic information,which is used as the input of the fully connected layer for classification.The results show that multi-scale features extraction method can better extract features of different scales compared with the single-scale feature extraction method,under the situation of complex ground object scene,which remarkably improves the classification performance of HSIs.In order to solve that it is difficult to obtain better classification results for hyperspectral image classification when the labeled training samples are insufficient,a hyperspectral image classification method based on deep model transfer is proposed.For which the source domain and target domain samples are both hyperspectral data in model based transfer learning,the source domain dataset is divided into samples of different proportions and then pre-trained on the deep neural network.Using a few number of labeled target domain samples to fine-tune the model and extract the specific feature in the target domain,and the spatial and spectral feature is combined in the hyperspectral image.In the case of insufficient labeled training data,transfer learning between heterogeneous data can significantly improve the classification results.Aiming at the transfer task between natural images and hyperspectral images,and fully seeking the potential of existing natural image data sets in model transfer,an asymmetric convolutional transfer learning(ACTL)model is proposed with replace the convolution kernel commonly used in deep neural networks with an asymmetric convolution block,which increase the weight of the central area of the convolution kernel and improve the feature extraction ability of the existing model.Maximum mean difference is used as the loss function of the model to reduce the difference between different dataset.Experiment result shows that ACTL effectively improves the classification effect when the training samples are less labeled and provides new ideas for transfer learning between heterogeneous dataset.