One-Class Risk Estimation for One-Class Hyperspectral Image Classification

TL;DR

This work tackles hyperspectral one-class classification by introducing HOneCls, a weakly supervised deep classifier that learns from positive and unlabeled data. Central to the approach is the One-Class Risk Estimator, which combines a conformance-based negative risk with a Positive Representation Enhancement, including a distribution-rebalancing parameter $\alpha_p$ and a focal-like positive risk controlled by $\gamma$, to address distribution overlap and imbalance. The estimator is backed by a consistency proof and practical training strategies (warm-up with BCE, sigmoid loss, and probability clamping) and is implemented in FreeOCNet, a lightweight FCN that preserves spectral-spatial details. Empirical results on HongHu, LongKou, and HanChuan show substantial gains over baselines and existing HSI target detection methods, validating the method’s robustness to few-shot positive data, distribution imbalance, and domain shift, with clear implications for reducing annotation burden in fine-grained HSI tasks.

Abstract

Hyperspectral imagery (HSI) one-class classification is aimed at identifying a single target class from the HSI by using only knowing positive data, which can significantly reduce the requirements for annotation. However, when one-class classification meets HSI, it is difficult for classifiers to find a balance between the overfitting and underfitting of positive data due to the problems of distribution overlap and distribution imbalance. Although deep learning-based methods are currently the mainstream to overcome distribution overlap in HSI multiclassification, few studies focus on deep learning-based HSI one-class classification. In this article, a weakly supervised deep HSI one-class classifier, namely, HOneCls, is proposed, where a risk estimator,the one-class risk estimator, is particularly introduced to make the fully convolutional neural network (FCN) with the ability of one class classification in the case of distribution imbalance. Extensive experiments (20 tasks in total) were conducted to demonstrate the superiority of the proposed classifier.

Figures (10)

• Figure 1: Characteristics of the HSI one-class classification. (a) Distribution overlap. Take the FanCun multi-spectral dataset 7486129 (positive class: tree) and HongHu HSI dataset (positive class: tuber mustard) as example. Using tSNE to display the data, the overlap between the distribution of positive and negative HSI data is more severe due to the negative HSI data contain other plants species. (b) Distribution imbalance. Take the HongHu dataset as an example, the classes of objects and their class probability are shown, the percentage of more than 11 kinds of objects is less than 0.05. (c) Instance of risk curves of distribution balanced and imbalanced data during training stage (the number of positive training samples for cotton and carrot is fixed at 100)
• Figure 2: Flowchart of the proposed HOneCls, which includes the module of global spectral-spatial features extractor and the module of one-class representation learning. One-Class Risk Estimator is responsible for training neural networks to extract robust spectral-spatial features to overcome the problem of distribution overlap in the case of distribution imbalance in one-class classification.
• Figure 3: Hyperspectral imagery with ground truth. (a) HongHu Dataset. (b) LongKou dataset. (c) HanChuan dataset.
• Figure 4: Distribution maps for the HongHu dataset.
• Figure 5: Distribution maps for the LongKou dataset.