Perceive, Act and Correct: Confidence Is Not Enough for Hyperspectral Classification

TL;DR

Hyperspectral image classification often misinterprets high predictive confidence as correctness, neglecting epistemic uncertainty. CABIN introduces a cognitive-aware semi-supervised framework that forms a closed loop of perception, action, and correction: perception estimates uncertainty with $\text{EDL}$, action uses $UGDSS$ to explore uncertain regions while preserving reliable pseudo-labels, and correction employs $FDAS$ guided by the Uncertainty-Gap metric $UG_\alpha$ to split pseudo-labels into reliable, ambiguous, and noisy sets with tailored losses. The approach uses $\mathcal{D}_t$, $\mathcal{D}_{hc}$, $\mathcal{D}_{qu}$, $\mathcal{D}_{au}$, $\widehat{\mathcal{D}}_{aug}$, $\mathcal{D}_{re}$, $\mathcal{D}_{am}$, and $\mathcal{D}_{no}$ within a unified loss $\mathcal{L} = \mathcal{L}_{\text{EDL}}(\mathcal{D}_L \cup \widehat{\mathcal{D}}_{aug}) + \lambda_r \mathcal{L}_{\text{EDL}}(\mathcal{D}_{re}) + \lambda_a \mathcal{L}_{\text{GCE}}(\mathcal{D}_{am})$. Empirically, CABIN yields consistent gains across five datasets and multiple backbones, improves labeling efficiency (e.g., effective with 50–75% annotations), and demonstrates robust, model-agnostic applicability in semi-supervised HSI classification. This work advances practical hyperspectral analysis by turning uncertainty into an active driver for data selection and supervision correction, reducing annotation costs while enhancing generalization.

Abstract

Confidence alone is often misleading in hyperspectral image classification, as models tend to mistake high predictive scores for correctness while lacking awareness of uncertainty. This leads to confirmation bias, especially under sparse annotations or class imbalance, where models overfit confident errors and fail to generalize. We propose CABIN (Cognitive-Aware Behavior-Informed learNing), a semi-supervised framework that addresses this limitation through a closed-loop learning process of perception, action, and correction. CABIN first develops perceptual awareness by estimating epistemic uncertainty, identifying ambiguous regions where errors are likely to occur. It then acts by adopting an Uncertainty-Guided Dual Sampling Strategy, selecting uncertain samples for exploration while anchoring confident ones as stable pseudo-labels to reduce bias. To correct noisy supervision, CABIN introduces a Fine-Grained Dynamic Assignment Strategy that categorizes pseudo-labeled data into reliable, ambiguous, and noisy subsets, applying tailored losses to enhance generalization. Experimental results show that a wide range of state-of-the-art methods benefit from the integration of CABIN, with improved labeling efficiency and performance.

Figures (12)

• Figure 1: Existing methods depend on confidence for both sampling and pseudo-labeling, yet overlook cognitive gaps, supporting our core insight that confidence is not enough.
• Figure 2: An overview of the proposed CABIN framework, designed to address the challenge that "confidence is not enough" for HSI classification. CABIN establishes a closed-loop process of perception, action and correction: the model first perceives sample uncertainty using EDL, then acts by selecting key samples via UDGSS, and finally corrects cognitive-behavioral gaps by dynamically assessing pseudo-label reliability with FDAS. This process enables more robust learning under limited supervision.
• Figure 3: Schematic of the UGDSS module, which leverages epistemic uncertainty to orchestrate a dual-path selection strategy balancing exploration and exploitation. The exploration path combines DRQS and GFP to actively probe the model's knowledge gaps, whereas the exploitation path selects confident samples for reliable pseudo-labeling.
• Figure 4: Left: Comparison of sampling strategies. (a) uses only dynamic thresholding, whereas (b) applies DRQS for diverse and non-redundant selection in feature space. Green circles: selected samples; red crosses: initial labeled samples. Right: Evolution of UG’s ability to discriminate sample difficulty throughout training.
• Figure 5: t-SNE projection of the original and augmented samples in feature space. Each color represents a semantic class. Light-colored points indicate original samples; dark points with outlines are selected samples; dark points without outlines are generated samples. Light and dark tones of the same color correspond to the same class.