top of page

Unsupervised Representation Learning for Leaf Trait Prediction and Spectral Novelty Detection Across 8 EcoSIS Datasets

Vidyut Krishnan
21/11/26

Non-destructive monitoring of plant traits through leaf-level reflectance spectroscopy is essential for understanding the functioning of global ecosystems and biodiversity across different biomes. Nevertheless, there are difficulties in combining results of diverse field experiments because of differences in hardware, spectral resolution, and sampling methods. Most workflows do not enable creation of coherent representations without extensive annotation of traits in diverse plant communities. In order to solve this problem, we have combined 8 EcoSIS datasets representing both temperate and tropical communities by bringing 13,549 leaf-level spectra to 201 consistent bands (400-2400 nm) by means of Savitzky-Golay smoothing and SNV normalization and then trained an unsupervised deep autoencoder to obtain 192-dimensional embeddings. When evaluating representations using group-based cross-validation in order to avoid taxa leakage, we found out that raw spectra yielded the best results for structural and water traits (R2 up to 0.92 for leaf mass per area and equivalent water thickness), whereas 192-D embeddings were comparable or better for estimating pigments (R2 = 0.94 - 0.96 on the Angers dataset). In the third stage, the latent space allowed us to perform the discovery task that enables the estimation of species-level spectral novelty based on scan frequency. This discovery task ranked Carex Richardsonii as the most spectrally novel species, with a score of 0.9525.

bottom of page