Remote sensing-based high-resolution mapping of the forest canopy height: some models are useful, but might they be even more if combined?

Abstract. The development of high-resolution mapping models for forest attributes based on remote sensing data combined with machine or deep learning techniques, has become a prominent topic in the field of forest observation and monitoring. This has resulted in an extensive availability of multiple sources of information, which can either lead to a potential confusion, or to the possibility to learn both about models and about forest attributes through the joint interpretation of multiple models. This article seeks to endorse the latter, by relying on the Bayesian model averaging (BMA) approach, which can be used to diagnose and interpret differences among predictions of different models. The predictions in our case are the forest canopy height estimations for the metropolitan France coming from five different models (Lang et al., 2023; Liu et al., 2023; Morin et al., 2022; Potapov et al., 2021; Schwartz et al., 2024). An independent reference dataset, containing four different definitions of the forest height (dominant, mean, maximum and Lorey’s), comes from the French National Forest Inventory (NFI) providing some 5 500 plots used in the study, distributed across the entire area of interest. In this contribution, we line up the evoked models with respect to their probabilities to be the ones generating measurements/estimations at the NFI plots. Stratifying the probabilities based on French sylvo-ecological regions reveals spatial variation in the respective model probabilities across the area of interest. Furthermore, we observe significant variability in these probabilities depending on the forest height definition used. This leads us to infer that the different models inadvertently exhibit dominant predictions for different types of canopy height. We also present the respective inter-model and intra-model variance estimations, allowing us to come to understand where the employed models have comparable weights but contrasted predictions. We show that the mountainous terrain has an important impact on the models spread. Moreover, we observe that the forest stand vertical structure, the dominant tree species and the type of forest ownership systematically appear to be statistically significant factors influencing the models divergence. Finally, we demonstrate that the derived mixture models exhibit higher R² scores and lower RMSE values compared to individual models, although they may not necessarily exhibit lower biases.