Physically Based, Machine Learning and Hybrid Approaches for Improved Plant Traits Retrievals from Future CHIME Hyperspectral Imagery

The forthcoming availability of new hyperspectral imaging systems on-board satellites will soon boost the potential of Earth Observation data in the vegetation domain. Among the on-going international initiatives, the CHIME mission of the European Space Agency is expected to expand the Copernicus constellation providing routine hyperspectral imagery in the 400-2500 nm spectral domain. Such data are critical for the quantitative estimation of plant traits (PTs) across space and time, which is essential to assess the dynamic response of vegetation. The contiguous spectral information carried by the hyperspectral data in fact may provide the means to obtain more accurate retrievals, but other aspects such as the robustness, efficiency and speed of the retrieval approach employed are still underevaluated. Within this framework, we examined the potential of the CHIME satellite for the operational mapping of leaf chlorophyll content (LCC) and leaf area index (LAI) testing different retrieval approaches. To explore this capability, we exploited hyperspectral airborne images collected with the HyPlant sensor (SPECIM Ltd., Finland) over a corn field located in Grosseto (Italy) in July 2018. The reflectance images were resampled to CHIME-like spectral and spatial configuration and used as input of three retrieval approaches: i) look-up-table based inversion of radiative transfer models (RTM); ii) machine learning regression algorithms (MLRA) and iii) hybrid approaches combining RTM simulations with MLRA. The results were compared against LCC and LAI ground validation data, collected in the field in 88 plots of 10 × 10 m2. Results in cross-validation showed a high performance of both RTM and MLRA in the retrieval of LAI (RMSE=0.46 m2 m-2, nRMSE=8.3%, r2=0.89 and RMSE=0.48 m2 m-2, nRMSE=8.7%, r2=0.89, respectively). Conversely, the MLRA outperformed the RTM in the retrieval of LCC (RMSE=4.01 µg cm-2, nRMSE=16.9%, r2=0.58 and RMSE=6.41 µg cm-2, nRMSE=24.2%, r2=0.21, respectively). These results, together with the possibility to provide retrieval uncertainties and the overall lower computational demand, highlighted the potential of MLRA as a promising tool for the operational mapping of PTs from next generation satellites.