A Survey: Deep Learning for Hyperspectral Image Classification with Few Labeled Samples (2112.01800v1)

Abstract: With the rapid development of deep learning technology and improvement in computing capability, deep learning has been widely used in the field of hyperspectral image (HSI) classification. In general, deep learning models often contain many trainable parameters and require a massive number of labeled samples to achieve optimal performance. However, in regard to HSI classification, a large number of labeled samples is generally difficult to acquire due to the difficulty and time-consuming nature of manual labeling. Therefore, many research works focus on building a deep learning model for HSI classification with few labeled samples. In this article, we concentrate on this topic and provide a systematic review of the relevant literature. Specifically, the contributions of this paper are twofold. First, the research progress of related methods is categorized according to the learning paradigm, including transfer learning, active learning and few-shot learning. Second, a number of experiments with various state-of-the-art approaches has been carried out, and the results are summarized to reveal the potential research directions. More importantly, it is notable that although there is a vast gap between deep learning models (that usually need sufficient labeled samples) and the HSI scenario with few labeled samples, the issues of small-sample sets can be well characterized by fusion of deep learning methods and related techniques, such as transfer learning and a lightweight model. For reproducibility, the source codes of the methods assessed in the paper can be found at https://github.com/ShuGuoJ/HSI-Classification.git.

• The paper analyzes transfer, active, and few-shot learning paradigms to address the challenge of classifying hyperspectral images with scarce labeled data.
• It demonstrates that models like relation networks and CNN_HSI can achieve robust classification performance even with minimal training examples.
• The study highlights the promise of lightweight architectures and graph-based approaches as future directions for enhancing hyperspectral image analysis.

Deep Learning for Hyperspectral Image Classification with Limited Labeled Samples: A Survey

The paper "A Survey: Deep Learning for Hyperspectral Image Classification with Few Labeled Samples" provides a comprehensive examination of current methodologies for hyperspectral image (HSI) classification, with a specific focus on scenarios characterized by limited labeled data. The authors highlight how hyperspectral images, rich in both spectral and spatial information, offer precise pixel-level classification capabilities that are crucial in various applications such as precision agriculture, environmental protection, and medical diagnostics. Given the significant hurdles associated with obtaining labeled hyperspectral images, the paper systematically addresses deep learning approaches that mitigate these challenges using limited training data.

Research Focus

The paper is structured around the exploration of different learning paradigms suitable for tackling the lack of labeled hyperspectral data. The authors segment the research into three primary categories of learning paradigms: transfer learning, active learning, and few-shot learning. Each paradigm is thoroughly reviewed to evaluate its applicability and effectiveness in the context of HSI classification.

1. Transfer Learning: The authors discuss fine-tuning methodologies and domain adaptation strategies. By leveraging pre-trained models on abundant source domain data and adapting them to a target domain with few samples, transfer learning provides a pragmatic approach to bridging the gap between extensive data requirements of deep learning models and limited availability of labeled hyperspectral data.
2. Active Learning: This paradigm intelligently selects sample queries to be labeled, thus maximizing the information gain from each labeled instance. The authors provide insights into committee-based and posterior probability-based methods for active learning, demonstrating how these can significantly enhance model performance on limited data.
3. Few-shot Learning: As a meta-learning paradigm, few-shot learning focuses on learning the differentiation between classes from a small number of examples. This section of the paper examines metric-based models such as the prototype network and relation network, detailing their adaptability to hyperspectral data classification.

Experimental Validation

To empirically validate the efficacy of these approaches, the paper includes comprehensive experiments on three well-known HSI datasets: PaviaU, Salinas, and KSC. Several state-of-the-art models, including autoencoders, convolutional neural networks (CNNs), and recurrent neural networks (RNNs) tailored for spectral-spatial feature extraction, were tested under configurations with different sample sizes per class (10, 50, 100). Results suggest that methods like the relation network (S-DMM) and CNN-based techniques like CNN_HSI display robustness even with minimal labeled data, underscoring the potential of metric-based learning approaches and lightweight network architectures in HSI classification problems.

Implications and Future Directions

The survey highlights key findings on lightweight architectures, the potential of combining different learning paradigms, and the adaptability of RNNs and transformers in hyperspectral data contexts. The integration of graph convolutional networks demonstrates emerging interest and effectiveness in hyperspectral image analysis, catering to non-Euclidean data structures typical in HSIs.

Conclusion

By offering substantial experimental evidence and exploration of various learning paradigms, the paper contributes to advancing the field of deep learning for HSI classification in contexts where labeled data is scarce. The synthesis of current methodologies and their respective challenges provides a bedrock for future research, with highlighted prospects in model lightweighting, the application of transformer architectures, and the deployment of graph-based learning methods for hyperspectral data.

Overall, the paper serves as a critical resource for researchers working in hyperspectral image analysis, emphasizing the importance of innovative deep learning solutions that accommodate the constraints of limited labeled data while maximizing classification efficacy. The potential continuous advancements in these areas pave the way for more adaptive, accurate, and scalable applications across various domains reliant on hyperspectral imaging technologies.