Deep learning has achieved great success in hyperspectral image classification. However, when processing new hyperspectral images, the existing deep learning models must be retrained from scratch with sufficient samples, which is inefficient and undesirable in practical tasks. This paper aims to explore how to accurately classify new hyperspectral images with only a few labeled samples, i.e., the hyperspectral images few-shot classification. Specifically, we design a new deep classification model based on relational network and train it with the idea of meta-learning. Firstly, the feature learning module and the relation learning module of the model can make full use of the spatial–spectral information in hyperspectral images and carry out relation learning by comparing the similarity between samples. Secondly, the task-based learning strategy can enable the model to continuously enhance its ability to learn how to learn with a large number of tasks randomly generated from different data sets. Benefitting from the above two points, the proposed method has excellent generalization ability and can obtain satisfactory classification results with only a few labeled samples. In order to verify the performance of the proposed method, experiments were carried out on three public data sets. The results indicate that the proposed method can achieve better classification results than the traditional semisupervised support vector machine and semisupervised deep learning models.

Satellite hyperspectral remote sensing has gradually become an important means of Earth observation, but the existence of various types of noise seriously limits the application value of satellite hyperspectral images. With the continuous development of deep learning technology, breakthroughs have been made in improving hyperspectral image denoising algorithms based on supervised learning; however, these methods usually require a large number of clean/noisy training pairs, a target that is difficult to meet for real satellite hyperspectral imagery. In this paper, we propose a self-supervised learning-based algorithm, 3S-HSID, for denoising real satellite hyperspectral images without requiring external data support. The 3S-HSID framework can perform robust denoising of a single satellite hyperspectral image in all bands simultaneously. It first conducts a Bernoulli sampling of the input data, then uses the Bernoulli sampling results to construct the training pairs. Furthermore, the global spectral consistency and minimum local variance are used in the loss function to train the network. We use the training model to predict different Bernoulli sampling results, and the average of multiple predicted values is used as the denoising result. To prevent overfitting, we adopt a dropout strategy during training and testing. The results of denoising experiments on the simulated hyperspectral data show that the denoising performance of 3S-HSID is better than most state-of-the-art algorithms, especially in terms of maintaining the spectral characteristics of hyperspectral images. The denoising results for different types of real satellite hyperspectral data also demonstrate the reliability of the proposed method. The 3S-HSID framework provides a new technical means for real satellite hyperspectral image preprocessing.

Albeit hyperspectral image (HSI) classification methods based on deep learning have presented high accuracy in supervised classification, these traditional methods required quite a few labeled samples for parameter optimization. When processing HSIs, however, artificially labeled samples are always insufficient, and class imbalance in limited samples is inevitable. This study proposed a Transformer-based framework of spatial–spectral–associative contrastive learning classification methods to extract both spatial and spectral features of HSIs by the self-supervised method. Firstly, the label information required for contrastive learning is generated by a spatial–spectral augmentation transform and image entropy. Then, the spatial and spectral Transformer modules are used to learn the high-level semantic features of the spatial domain and the spectral domain, respectively, from which the cross-domain features are fused by associative optimization. Finally, we design a classifier based on the Transformer. The invariant features distinguished from spatial–spectral properties are used in the classification of satellite HSIs to further extract the discriminant features between different pixels, and the class intersection over union is imported into the loss function to avoid the classification collapse caused by class imbalance. Conducting experiments on two satellite HSI datasets, this study verified the classification performance of the model. The results showed that the self-supervised contrastive learning model can extract effective features for classification, and the classification generated from this model is more accurate compared with that of the supervised deep learning model, especially in the average accuracy of the various classifications.

Accurate image classification under extreme label scarcity remains a critical challenge in remote sensing, where hyperspectral imagery offers rich spectral signatures but suffers from coarse spatial resolution limitations. Although integrating complementary panchromatic and multispectral data can alleviate this spatial bottleneck, existing methods struggle to fully exploit their deep spatial–spectral features without dense annotations. We propose GeCoM, a joint generative and interactive contrastive self-supervised learning framework for few-shot multi-source image classification. GeCoM employs modality-specific Transformer encoders with a multi-scale masking strategy to reconstruct fine spatial–spectral patterns. Simultaneously, an interactive contrastive optimization module is introduced to align heterogeneous multi-source features within a unified latent space. Experiments on the Houston 2018 benchmark and two regional scenes from Gaofen-14 satellite demonstrate that GeCoM achieves superior Overall Accuracies of 74.66%, 92.36%, and 94.02% with only 15 labeled samples per class, outperforming state-of-the-art baselines. This study provides a robust, label-efficient spatial–spectral representation solution for multi-sensor earth observation.