Prospects for Mitigating Spectral Variability in Tropical Species Classification Using Self-Supervised Learning

TL;DR

This paper addresses the challenge of spectral variability in tropical forest species classification from airborne hyperspectral data. It introduces a self-supervised learning framework based on the Barlow-Twins objective to learn robust spectral representations from inter-date pixel pairs, with the loss $\mathcal{L} = \sum_{k}(1 - C_{kk})^{2} + \lambda \sum_{k}\sum_{l \neq k} C_{kl}^{2}$ guiding redundancy minimization across the cross-correlation matrix $C$ of projected features $Z_{T1}$ and $Z_{T2}$. A downstream Linear Discriminant Analysis classifier evaluates cross-date robustness on 40 species, demonstrating about $10$ points higher accuracy stability compared to reflectance-based features. The approach highlights the potential of SSL with temporally informed views to produce more stable spectral descriptors for tropical forest monitoring, while acknowledging the need for broader geographic validation and acquisition alignment in future work.

Abstract

Airborne hyperspectral imaging is a promising method for identifying tropical species, but spectral variability between acquisitions hinders consistent results. This paper proposes using Self-Supervised Learning (SSL) to encode spectral features that are robust to abiotic variability and relevant for species identification. By employing the state-of-the-art Barlow-Twins approach on repeated spectral acquisitions, we demonstrate the ability to develop stable features. For the classification of 40 tropical species, experiments show that these features can outperform typical reflectance products in terms of robustness to spectral variability by 10 points of accuracy across dates.

Figures (3)

• Figure 1: Schematics of different physical process such as atmospheric variability, sun and view angle variability over a mapped coordinate of a canopy.
• Figure 2: Overall workflow for the Self-Supervised training across acquisitions.
• Figure 3: Test dataset mean classification accuracy for 40 tropical species for each SSL pre-training strategy compared to LDA results on pre-processed reflectance. For each data-augmentation strategy (x-axis), red bars show accuracy with the positive pair construction strategy, and blue bars without pairing. Bars indicate mean accuracy across all runs, with uncertainty intervals showing standard deviation from four independent runs. Dotted grey and black lines mark the average classification accuracy on the training and test datasets, respectively.