Gaussian Processes Regression and PLSR for mapping forest canopy traits from Fenix Airborne Hyperspectral Data

Rui Xie, R. Darvishzadeh, A.K. Skidmore, Marco Heurich, Stefanie Holzwarth, Tawanda Gara, Ils Reusen

Research output: Contribution to conferenceAbstractAcademic

48 Downloads (Pure)


Machine learning algorithms, and specifically kernel-based methods such as Gaussian processes regression (GPR), have been shown to outperform traditional empirical methods for retrieving vegetation traits. GPR is attractive for its property of automatically generating uncertainty estimates for predicted traits. GPR has been increasingly used for the estimation of canopy traits from hyperspectral remote sensing data in agricultural fields and grassland ecosystems. However, to our knowledge, the application of GPR using full-spectrum airborne hyperspectral data in forest ecosystems remains under-explored. Therefore, in this study, we evaluated the performance of GPR as a representative of kernel-based machine learning algorithms in estimating two essential forest canopy traits (i.e., LAI and canopy chlorophyll content) using airborne hyperspectral data. The performance of GPR was compared with partial least square regression (PLSR) which is widely used for retrieving vegetation traits in spectroscopic studies. Field measurements of LAI and leaf chlorophyll content were collected in the Bavarian Forest National Park (BFNP) in Germany, concurrent with the acquisition of the Fenix airborne hyperspectral data (400−2500 nm) in July 2017 in the framework of the EUFAR summer school RS4forestEBV. The cross-validated coefficient of determination (R2) and normalised root mean square error (nRMSE) between the field-measured and retrieved traits were used to examine the accuracy of the respective methods. The results indicated that GPR somewhat outperformed PLSR in producing accurate estimations for LAI (GRP nRMSE = 16.7%; PLSR nRMSE = 23.0%) and canopy chlorophyll content (GPR nRMSE = 16.2%; PLSR nRMSE = 22.5%). The uncertainty maps generated by GPR showed that the retrieval uncertainties were generally low across the map, whereas higher uncertainties mainly corresponded with regions with low vegetation cover or under-represented in our field sampling. The capability to generate accurate predictions and associated uncertainty estimates suggest the GPR may be a promising candidate for the retrieval of vegetation traits.
Original languageEnglish
Number of pages1
Publication statusPublished - 21 Jun 2022
Event12th EARSeL Workshop on Imaging Spectroscopy 2022 - Potsdam, Germany, Potsdam, Germany
Duration: 22 Jun 202224 Jun 2022
Conference number: 12


Conference12th EARSeL Workshop on Imaging Spectroscopy 2022
Internet address


Dive into the research topics of 'Gaussian Processes Regression and PLSR for mapping forest canopy traits from Fenix Airborne Hyperspectral Data'. Together they form a unique fingerprint.

Cite this