Optical Remote Sensing Images Recognition By Means Of Few-shot Learning

The retrieval and classification of the high spatial resolution remote sensing image(HSR-RSI)and the hyperspectral image(HSI)are important research issues in optical remote sensing image recognition.Recently,the rapid development of deep learning algorithms has brought opportunities to identify HSR-RSI and HSIs efficiently.Many deep learning algorithms usually require a large number of labeled samples.However,the manual generation of tags is time-consuming and labor-intensive.Therefore,the new technologies and methods of the few-shot learning need to be developed.From the related algorithms of deep learning,with the data sets of high-resolution remote sensing images(HSR-RSI)and hyperspectral image(HSI),our work mainly focuses on the image retrieval and image classification using few-shot learning.The main work and conclusions are as follows:(1)Based on the few-sample learning,we propose an HSR-RSI retrieval method.We first employ high-level feature extraction(HFE)to extract high-level features.Then,the high-level features obtained by HFE are provided for DML.The multilayer DML is used to increase the intraclass compactness and interclass separability for HSR-RSI retrieval.Moreover,the extracted features are further processed with GAN to mitigate the overfitting problem.In the GAN,the generator synthesizes fake HSR-RSI similar to real HSR-RSI.In the experiment,our method is compared with the related approaches.The experimental results on the three data sets demonstrate the proposed method's superior performance over state-of-the-art techniques in HSR-RSI retrieval.Our method obtains improvements of 11.88%,18.19%,and 23.91% in m AP%compared to the related approaches with 2% labeled samples.(2)Based on the subspace learning(SL),we propose an HSI classification method.Combining a small amount of labeled data with a certain unlabeled data,the CRF is embedded into the subspace learning framework.First,a 3DCAE is trained to obtain the representation of the latent subspace.The relationship matrix is further formulated by combining the relationships of the latent space and the spatial distance.The CRF unary and pairwise terms are constructed using the probabilities of pixels belonging to different categories(only using a small number of labeled samples)and the relationships between the pixels.SL and relationship matrix can learn an accurate representation of subspace without using the labeled samples,which reduces the dependence of the labeled samples.The developed method is tested using three public adopted HSI data sets.The results can validate that the proposed method can perform better than the related HSI classification approaches.Our method obtains improvements of 11.08%,4.21%,and5.26% in OA% compared to the related approaches.(3)Based on feature consistency,we propose an HSI classification method.Without the labeled samples,feature consistency constraints the single pixel and group pixels,respectively.First,with CNN feature extractor,we extract spectral-spatial features from a 3-D patch.At the same time,fully connected layers(FCLs)model the feature consistency,including the feature consistency of single pixel(FCS)and feature consistency of group pixels(FCG).The FCS is achieved by the generative adversarial network(GAN)regularization,which can reconstruct the original data from extracted features.The FCG is based on the assumption that the features of group pixels should have similar characteristics within a superpixel,which is embedded in each FCL.The final FCL outputs the class labels,and the cross-entropy loss(CEL)is calculated with the labeled samples to boost the classification performance.The results tested on three HSI data sets can validate that the proposed method outperforms the related state-of-the-art HSI classification methods.Our method obtains improvements of 4.41%,3.20%,and 4.43% in OA% compared to the associated approaches.