Research Under Sample Constrained Conditions Methods For Hyperspectral Remote Sensing Image Classification

With the massive growth of hyperspectral data,how to effectively and accurately identify features on hyperspectral images has become one of the hot research problems in the field of machine learning.However,the high spectral variability of hyperspectral images makes visual interpretation difficult,and the acquisition of marker samples is more difficult compared with natural images and high-resolution images.Therefore,the problem of limited samples is more severe in the task of hyperspectral image classification.It can be said that the problem of limited samples at this stage greatly hinders the pace of application of hyperspectral image classification and affects the interpretation accuracy of images.Therefore,how to classify hyperspectral images under sample-constrained conditions and reduce the reliance on labeled samples has become a frontier scientific problem in the field of remote sensing.The sample-constrained problem in hyperspectral classification tasks can be summarized as:(1)The problem of noisy samples: a large number of labeled samples are available,but it is difficult to guarantee the accuracy of their labels;(2)The number of labeled samples appears to be too small for the high dimensionality of hyperspectral data,posing the small sample problem;(3)In the extreme case,the number of labeled samples is zero,which is regarded as the no-sample problem.Although some machine learning methods have been proposed to overcome the above sample-constrained problems,the following problems still exist:(1)Existing hyperspectral image noise sample learning methods obtain classification results through pre-processing steps such as noise label detection and purification,combined with traditional classifiers(e.g.,support vector machines).How to directly learn deep classifiers with robustness to noisy samples remains a difficult task;(2)Feature dimensionality reduction is one of the means to solve the problem of small samples for hyperspectral classification tasks;however,traditional feature dimensionality reduction methods cannot effectively capture spatial correlation and spectral-spatial features are not sufficiently mined;(3)Existing unsupervised deep feature extraction methods focus on the underlying data structure,and the high-level semantic mining is inadequate to meet the needs of advanced semantic classification tasks.To address the above issues,this paper conducts a systematic study of hyperspectral classification under the condition of limited samples.The main research contents include:(1)The solutions to the problem of sample limitation in hyperspectral classification tasks are systematically summarized,and their shortcomings are discussed and analyzed in detail.(2)In terms of noisy sample learning,a sample selection network(S2Net)is proposed for collaborative hyperspectral image classification to address the problem that noisy samples lead to significant performance degradation of deep learning networks.The method designs a novel network training framework and uses a crossselection strategy to reduce the cumulative error caused by sample selection bias.On the hyperspectral data disturbed by noise samples,this method can ensure superior object recognition accuracy,and can be transplanted into the existing advanced deep network model,which satisfies the generalization and reliability requirements of the hyperspectral image classification.(3)In the aspect of small sample learning,feature dimensionality reduction is one of the effective methods to solve the small sample problem.Aiming at the limitation of traditional feature dimension reduction methods in mining the spatial correlation of data,a spatial spectral feature dimension reduction algorithm(SSDR)is developed.By constructing a superpixel-guided graph,the spatial correlation of the data after feature dimensionality reduction is adaptively constrained by effectively exploiting to the large amount of unlabeled sample information present in the image.At the same time,a few label-guided graph are constructed and multi-feature fusion is achieved.SSDR not only significantly reduces the computational burden,but also improves the classification accuracy under small samples.(4)In terms of unsampled feature learning,we propose a spectral-spatial semantic feature learning network(S3FN)to address the problem that existing unsupervised deep feature extraction methods perform feature extraction through data reconstruction and insufficient high-level semantic mining.With the proposed high-level semantic alignment strategy,S3 FN can learn discriminative high-level semantic spectral-spatial features,which are more suitable for downstream semantic classification tasks.To enhance the efficiency of the network training process,an efficient spectral-spatial feature learning network is further designed to extract multi-level spectral and spatial features.S3 FN overcomes the problem of focusing on the underlying features caused by the existing data reconstruction-based feature extraction methods due to the limitation of the original data structure,and achieves the current state-of-the-art feature extraction performance.