Using synthetic data to evaluate the benefits of large field plots for forest biomass estimation with LiDAR

With the maturation of methods for estimating aboveground forest biomass by remote sensing, researchers increasingly need test data, particularly ground reference data, that are large enough to fine-tune existing approaches and test their robustness under diverse conditions. In this context, realistic synthetic datasets present an interesting alternative to costly and limited field data. Here, we present a new approach to simulate realistic canopy height and cover type data by combining an individual-tree forest simulator with real LiDAR point clouds of individual trees. We demonstrate the utility of our approach by re-examining the influence of field plot size on the predictive power of remote-sensing models for biomass estimation. Our approach with a complete (wall-to-wall) field reference dataset and matching synthetic remote sensing data allowed us to not only perform internal cross-validations with field plots that were used to fit the model (as in studies with real data), but to also consider the quality of model predictions to a standardized spatial grid or to the entire region. Our results confirm earlier reports of smaller predictive errors with increased field plot sizes under internal model validation (RMSE of 125 t/ha at 10 m field plot size and RMSE of 40 t/ha for 50 m field plots). However, we show that this is mainly an artifact of comparing the models with the same data they were fit, thus with validation data of different scales. When validating on a grid with standardized scale, smaller field plots performed almost equally well as larger field plots (small RMSE decrease between 4 t/ha and 7 t/ha when going from 10 m to 50 m plots), and even outperformed them if we assumed that increasing the plot size means that fewer field plots can be obtained (small RMSE increase between 9 t/ha and 12 t/ha when going from 10 m to 50 m plots). We conclude that synthetic remote sensing datasets are a useful tool for method testing. The suggested approach may be used to reexamine our current methodological understanding, which is often based on tests with real data of very limited sizes, as well as to optimize workflows, model choices and data collection. A wider use of synthetic data could be instrumental in improving remote sensing methodology.