Self-supervised learning and domain adaptation to address spectral variability in airborne hyperspectral acquisitions in tropical forests

Résumé

Tropical forests, which cover 6% of Earth's land surface, are essential for global biodiversity and carbon storage. In French Guiana, where forests account for over 97% of the territory, sustainable forest management represents a critical challenge for both conservation and economic development. The French National Forest Office (ONF) manages 6 million hectares, seeking to balance timber extraction with ecosystem preservation. Remote sensing provides a potential solution for large-scale tree species mapping, thereby supporting evidence-based forest management and reducing the environmental footprint of logging activities. This thesis investigates the challenge of generalizing tree species identification models using airborne hyperspectral imagery in hyperdiverse tropical environments. The study is structured around three scientific questions, analyzed through repeated acquisitions over the Paracou and Nouragues sites in French Guiana. The hyperspectral data, acquired in 2016, cover the 4002500 nm spectral range. First, the analysis quantifies the primary sources of spectral variability : atmospheric conditions, solar geometry, and viewing angles, and demonstrates their significant impact on reflectance stability, even after atmospheric correction. The results indicate that this variability disrupts spectral proxies commonly used in vegetation monitoring, calling into question their reliability under operational conditions. Second, the study evaluates the capacity of self-supervised learning (SSL) to generate spectral representations that are robust to acquisition-related variability. While SSL reduces abiotic noise, it may accentuate intra-individual and site-specific differences, thereby limiting its effectiveness for cross-site species classification. Third, the research assesses domain adaptation (DA) techniques to mitigate spectral shifts between dates and sites. The findings reveal that unsupervised DA methods yield only modest improvements, whereas supervised DA approaches, such as Transfert Learning Adaboost (TrAdaBoost), achieve substantial classification gains even with limited labeled data. However, their success remains contingent on the representativeness of annotated samples and the ecological complexity of the study sites. In conclusion, this work frames spectral variability as a central concept for addressing the complex challenge of transfer learning between study sites and acquisition dates in tropical forest remote sensing. The integration of SSL and DA methods provides a structured approach to managing this variability, enabling the development of more reliable and transferable species identification models in operationally relevant contexts.

Auteurs, date et publication :

Auteurs Colin Prieur

Date : 2025