Research On Hyperspectral Remote Sensing Image Classification Method Based On Few-shot Learning

Hyperspectral image classification based on deep learning has achieved great success,but it requires a large number of labeled samples,and the acquisition of labeled samples is very difficult,which costs lots of human,material and financial resources.In the real scene,new scene（target domain）images often have only few labeled samples,but other scene（source domain）images often have enough labeled samples.Therefore,the source scene with many labeled samples can be used to help the target scene with few labeled samples to classify.Fewshot learning is an effective method to realize few samples classification.It utilizes the idea of meta-learning,where general information learned from source domain data（meta-knowledge）can help make predictions about target domain data.To solve the issue of classification of target scenes with few labeled samples,this paper proposes two kinds of few-shot classification methods based on different source domain data sets.The main contents are as follows:（1）When hyperspectral data is used as the source domain data set,a deep cross-domain few-shot learning method is proposed,which solves the problem of few-shot learning and domain shift in a unified framework for the first time.Firstly,mapping is used to keep the channel same between domains.Then deep network is designed to extract deep spatial-spectral features,and conditional domain adaptive technology is used to reduce domain shift to achieve global domain distribution alignment.In addition,few-shot learning is performed on the source and target domain simultaneously,which can not only learn transferable knowledge from the source domain,but also learn a discriminative feature embedding model for the target domain.The effectiveness of this method has been proved on four different hyperspectral data sets.（2）When the natural image data is used as the source domain data set,a heterogeneous few-shot learning method is proposed to transfer the rich spatial information such as texture and structure from the natural image to the hyperspectral spatial feature extraction part to realize the transfer between the heterogeneous data sets.It uses the designed network to extract spectral features,and fuses the extracted spatial and spectral features to improve classification accuracy.Firstly,the mini-ImageNet data set is used for few-shot learning,and then the hyperspectral data set is used for fine tuning,which can not only transfer general knowledge learned from source domain,but also extract the robust spatial-spectral features of the target domain.The effectiveness of this method has been proved on three different hyperspectral data sets.