Self-Supervised Hyperspectral Learning for Leaf Multitrait Prediction With Sparse Labels

Accurate estimation of leaf traits is critical for evaluating ecosystem functions and elucidating vegetation–environment interactions. Hyperspectral remote sensing provides a non-destructive and information-rich means for trait analysis, yet its potential remains constrained by the scarcity of labeled data and the limited generalizability of existing learning methods. To overcome these challenges, we propose the Spectral Joint Embedding Predictive Architecture (S-JEPA), a self-supervised learning framework designed to learn generalizable spectral features by predicting target-region representations from contextual spectral information. Utilizing the largest unlabeled hyperspectral dataset to date, comprising nearly 100,000 samples, S-JEPA enables accurate trait estimation with minimal labeled data. Across six representative leaf traits, S-JEPA achieved an average R2of 0.784, substantially exceeding conventional supervised, transfer-learning, and multi-task learning models, and also surpassing another self-supervised baseline, S-MAE, by a small margin. Furthermore, S-JEPA demonstrate superior robustness especially in data-scarce scenarios, efficiently tackling a longstanding field challenge. The learned spectral features revealed coherent inter-trait associations aligned with known physiological linkages, demonstrating that S-JEPA captures biologically meaningful structures rather than statistical artifacts. This study highlights the efficacy of SSL in extracting generalizable spectral features and establishes a scalable framework for multi-trait estimation and ecosystem monitoring.