From spectra to plant functional traits: Transferable multi-trait models from heterogeneous and sparse data

• Published in: Remote Sensing of Environment
• Main Authors: Cherif, E., Feilhauer, Hannes, Berger, K., Dao, P.D., Ewald, M., Hank, T.B., He, Y., Kovach, K.R., Lu, B., Townsend, P.A., Kattenborn, T.
• Format: Article in Journal/Newspaper
• Language: English
• Published: Elsevier 2023
• Subjects: Hyperspectral remote sensing; Plant trait retrieval; Deep learning; Biophysical variables; Imaging spectroscopy; Canopy properties; Weakly supervised learning; Multi-task regression; Tundra
• Online Access: https://www.ufz.de/index.php?en=20939&ufzPublicationIdentifier=22445; https://doi.org/10.1016/j.rse.2023.113580

Large-scale information on several vegetation properties (‘plant traits’) is critical to assess ecosystem functioning, functional diversity and their role in the Earth system. Hyperspectral remote sensing of plant canopies offers a key tool to map multiple plant traits. However, we are still lacking generalized methods to translate hyperspectral reflectance into a suite of relevant plant traits across biomes, land cover and sensor types. The absence of globally representative data sets and the gap between the available reflectance data with corresponding in-situ measurements have hampered such approaches. In recent years, the scientific community acquired multiple data sets encompassing canopy hyperspectral reflectance and plant traits from different plant types and sensors. To combine these heterogeneous data sets, we propose three multi-trait modeling approaches based on Convolutional Neural Networks (CNNs) to simultaneously infer a broad set of 20 structural and chemical traits (e.g. leaf mass per area, leaf area index, pigments, nitrogen). The performance of these multi-trait CNN models predicting these traits was compared against single-trait CNN as well as single-trait partial least squares regression (PLSR). We found that the multi-trait CNNs performances significantly increased from single-trait CNNs (nRMSE = 0.027–19.61%) and the state-of-the-art PLSR models (nRMSE = 1.94–40.07%) across a broad range of vegetation types (crops, forest, tundra, grassland, shrubland) and sensor types. Thus, providing a single model for multiple traits not only proved to be computationally more efficient, but also more accurate, since it enabled the model to incorporate traits' co-variation. Despite the data heterogeneity of the merged data set, our models performances' were comparable or exceeded those of previous studies. Overall, this study highlights the potential of weakly supervised approaches to overcome the scarcity of in-situ measurements and take a step forward in creating efficient predictive models of multiple ...