Spectral-Spatial Contrastive Learning Framework for Regression on Hyperspectral Data

TL;DR

This work tackles regression on hyperspectral data using a spectral-spatial contrastive learning framework that is backbone-agnostic and compatible with both 3D-CNN and transformer architectures. It introduces a patch-based processing pipeline, a radius-based positive-pair strategy for regression targets, and a rich toolbox of spectral and spatial data augmentations, all optimized via a combined loss ${L_{total} = L_R + \alpha L_{Contrastive}}$. Empirical results on synthetic data and the Samson dataset demonstrate consistent improvements over baselines, with the spectral-spatial augmentation and transformer backbones delivering the strongest performance. The framework improves robustness to spectral and spatial variabilities and holds promise for practical hyperspectral regression tasks such as unmixing and pollution concentration prediction.

Abstract

Contrastive learning has demonstrated great success in representation learning, especially for image classification tasks. However, there is still a shortage in studies targeting regression tasks, and more specifically applications on hyperspectral data. In this paper, we propose a spectral-spatial contrastive learning framework for regression tasks for hyperspectral data, in a model-agnostic design allowing to enhance backbones such as 3D convolutional and transformer-based networks. Moreover, we provide a collection of transformations relevant for augmenting hyperspectral data. Experiments on synthetic and real datasets show that the proposed framework and transformations significantly improve the performance of all studied backbone models.

Figures (2)

• Figure 1: Architecture of the proposed framework.
• Figure 2: Original (blue) and transformed (orange) examples from real spectral data.