Remote-sensing-based forest canopy height mapping: Some models are useful, but might they provide us with even more insights when combined?

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 the availability of multiple, sometimes conflicting, sources of information, but, at face value, it also makes it possible to learn about forest attribute uncertainty through the joint interpretation of multiple models. This article seeks to endorse the latter by utilizing the Bayesian model averaging approach to diagnose and interpret the differences between predictions from different models. The predictions in our case are forest canopy height estimations for metropolitan France arising from five different models. An independent reference dataset, containing four different definitions of forest height (dominant, mean, maximum, and Lorey's) was established based on around 5500 plots of the French National Forest Inventory (NFI), distributed across the entire area of interest. In this study, we evaluate models with respect to their probabilities of correctly predicting measurements or estimations obtained from NFI plots, highlighting the spatial variability in respective model probabilities across the study area. We observed significant variability in these probabilities depending on the forest height definition used, implying that the different models inadvertently predict different types of canopy height. We also present the respective inter-model and intra-model variance estimations, enabling us to grasp where the employed models have comparable contributions but contrasting predictions. We show that topography has an important impact on the models spread. Moreover, we observed that the forest stand vertical structure, the dominant tree species, and the type of forest ownership systematically emerge as statistically significant factors influencing the model divergences. Finally, we observed that the fitted higher-order mixtures, which enabled the presented analyses, do not necessarily reduce bias or prevent the saturation of the predicted heights observed in the individual models.