Hyperspectral Image Classification Algorithm Based On Deep Learning And Few-Shot Learning

At present,with the rapid progress of the new round of technological revolution and industrial change,the proliferation of hyperspectral remote sensing images urgently requires the fine interpretation for tons of new fields.Although there are already various deep networks that are effective in hyperspectral interpretation tasks,the first requirement is to provide sufficient real-value data with high confidence to train and build models.In fact,for scenes with diverse landscapes and fine-grained complex terrain structures,relying on manual fine-grained annotation is a time-consuming and cost-effective process,and adequate annotation data are not available.The framework constructed due to insufficient sample size usually appear to be overfitting,which seriously hinders the decoding timeliness of practical tasks.In this paper,we focus on the task-driven detailed classification of hyperspectral images and combine the few-shot learning to construct classification networks,which can achieve accurate discrimination of images with a limited number of labeled samples.Most existing deep networks extract local spatial spectral information based on convolutional kernels,but ignore the potential relationships between non-local spatial samples.Such non-local inter-relationships can better characterize the statistical information of features when labeled samples are scarce.To this end,spatial attention and spectral query function modules are coupled in the feature extractor to overcome the constraints of the convolution kernel,which can consider association information between long-range(non-local)samples to improve the perceptual power of the framework.More significantly,a meta-learning strategy is employed to train the network,which can quickly implement the classification task in the target domain(TD)with the aid of the empirical knowledge of classification learned in the source domain(SD),relying on the supervision of a few marked instances.Due to the serious data distribution gap between the SD and TD,the models trained on the SD data are not applied to the TD directly,which seriously affects the effectiveness of the network.In this work,we design an adversarial domain discriminator by combining the idea of domain adaptation,which greatly shrinks the domain quality divergence and ensures the shared suitability of inter-domain meta-knowledge.In practical works,some sample elements in the image are multi-solvable by the potential influence of noise,and there may be some distortion in the matching of label information.For this reason,a hybrid attention mechanism is adopted to select purer samples for computing class prototypes.Meanwhile,to alleviate the negative effects caused by feature sparsity,more representative features are selected in the embedding space for metric classification.With respect to the existence of spatial-spectral shifts between domains,a maximum mean difference bias constraint between spatial and spectral is imposed on the depth semantic features,which in return realizes effective feature alignment under different degrees of spatial-spectral shifts,further enhancing the generalization power and decoding robustness of the model in the target domain.Extensive experiments on three publicly available hyperspectral datasets show that the proposed algorithm can produce better classification results relative to state-of-the-art algorithms.