Abstract
Hyperspectral images (HSIs) provide rich spectral and spatial information, yet their classification is often hindered by a lack of labeled training data. This paper presents two novel frameworks—OFSS-SL and MSS-MVSL—that integrate active learning with semi-supervised techniques. By employing secondary screening to select diverse and uncertain samples and a multiple-verification process for unlabeled data, these frameworks achieve superior classification accuracy with fewer labeled samples and reduced computational overhead compared to existing methods.
Introduction
Hyperspectral imaging captures hundreds of narrow spectral bands, offering a detailed "fingerprint" for every pixel. While this data is invaluable for remote sensing applications like agriculture and urban planning, the high dimensionality creates a "labeling bottleneck." Manual labeling by domain experts is expensive and time-consuming. To address this, researchers look toward Active Learning (AL), which identifies the most informative samples for labeling, and Semi-Supervised Learning (SSL), which leverages the vast amounts of available unlabeled data.
Proposed Frameworks
The authors propose two "syncretic" frameworks designed to maximize the utility of limited labeled data by merging the strengths of AL and SSL.
OFSS-SL: One-Fold Secondary Screening - SSL
The OFSS-SL framework introduces a "secondary screening" mechanism. Traditional active learning often selects samples based solely on uncertainty—points where the model is least confident. However, these points might be redundant if they are all similar. OFSS-SL clusters these uncertain samples to ensure the selected set is also diverse, representing different regions of the feature space.
MSS-MVSL: Multiple Secondary Screening - Multiple Verification SSL
MSS-MVSL is a more robust evolution of the concept. It utilizes multiple active learning strategies and clustering techniques simultaneously to build a "buffer pool" of candidate samples.
A key innovation in MSS-MVSL is the Multiple Verification step. To prevent the "poisoning" of the training set with incorrect pseudolabels—a common issue in SSL—the framework uses three distinct classifiers to vote on the labels of unlabeled pixels. A pseudolabel is only accepted if there is a consensus among the classifiers, ensuring high-confidence data augmentation.
The figure above illustrates the MCLU query function, a critical part of the uncertainty evaluation. It uses a One-Against-All (OAA) approach with multiple Support Vector Machines (SVMs) to determine which samples from the unlabeled pool are most informative for the model to learn next.
Experimental Results
The frameworks were rigorously tested on three benchmark datasets: Indian Pines, Kennedy Space Center (KSC), and Pavia University.
- Superior Accuracy: On the Indian Pines dataset, using only 200 labeled samples, the MSS-MVSL framework achieved an Overall Accuracy (OA) of 69.93%, outperforming the state-of-the-art CASSL method (66.89%).
- Efficiency: Similar performance gains were observed across the KSC and Pavia University datasets, where the proposed methods consistently reached high accuracy plateaus earlier in the labeling process than competing methods.
- Computational Cost: Despite the use of multiple classifiers for verification, the optimized screening process reduced the total training time. On the Indian Pines dataset, MSS-MVSL was approximately 16% faster than CASSL, demonstrating that smarter sample selection can lead to both better performance and lower costs.
Conclusion
The OFSS-SL and MSS-MVSL frameworks provide a powerful solution to the labeling bottleneck in hyperspectral image classification. By combining diversity-aware active learning with multi-model semi-supervised verification, these methods allow for high-accuracy classification with minimal human intervention. Future work will focus on scaling these frameworks for big data environments using distributed computing and exploring more diverse base classifiers to further enhance robustness.
Summary updated