Hybrid retrieval of canopy foliar biomass and nitrogen content between and beyond grassland ecosystems: trade-offs in ecosystem specificity versus model transferability

Accurate retrieval of biophysical variables is crucial for characterizing properties (i.e. traits) of plant canopies and capturing their spatiotemporal changes. Optical remote sensing offers the unique possibility for frequent and large-scale mapping of biophysical variables due to strong associations between spectral data and plant optical traits. One approach to formally predict plant properties remotely is based on hybrid retrieval. In this approach, a radiative transfer model (RTM) is used to simulate plant spectra for model training and then a machine learning regression is utilized for model prediction.

Hybrid retrieval approaches have two main advantages. First, the approach augments large field datasets needed for training with simulations modelled based on physical relationships between electromagnetic radiation and plant properties. The universal physics behind this have led to assumptions of greater transferability of these models when compared to empirical models. Second, the machine learning implementation provides the flexibility and computational efficiency of nonlinear nonparametric methods to link spectra and plant properties.

The recent implementation of active learning (AL) approaches offers promising and adaptive solutions to further enhance hybrid retrieval approaches. AL seeks to overcome the genericity and heavy assumptions of RTM simulations as opposed to the noisy real-world spectra and particularities of ecosystems by subsetting the training data to boost model performance. However, it is unclear how the selection of training data by an AL approach thereby affects model transferability and whether its selection relates to the ecology of different sites. Our work aims to assess how representative the AL-selected training samples are for their respective ecosystem and whether the generated models are transferable to other study sites.

Here, we used Gaussian process regression (GPR) trained with PROSAIL simulations in combination with AL to retrieve canopy foliar biomass and nitrogen content from Sentinel-2 data in three grassland sites with different characteristics, including alpine, prairie, and temperate grasslands in Switzerland, the United States, and Germany, respectively, and one heterogeneous forest and shrubland site in Portugal. We compared the trait space of the selected training samples with those of in-situ data and TRY database to assess their respective ecological representativeness. Further, we used our generated models to predict canopy foliar biomass and nitrogen across sites to check for their transferability.

Our preliminary results show promising accuracy of locally trained models to retrieve canopy foliar biomass (Switzerland: R² = 0.41, RMSE = 106.5 g/m²; United States: R² = 0.42, RMSE = 85.5 g/m²; Germany: R² = 0.28, RMSE = 96.2 g/m²; Portugal: R² = 0.6, RMSE = 60.9 g/m²). In particular, AL-selected training data increased model performances but was highly affected by the validation data thus limiting the general transferability of the models across study sites.

Based on these results, we can confirm adequate and stable performance of locally trained GPR-AL models. However, the transferability of such an approach requires further testing and an expanded search for solutions. For now, strong trade-offs exist between local optimization and transferability which challenges predictions of high accuracy across large spatial extents with limited field data.