Reconstructing the digital twin of forests from a 3D library: Quantifying trade-offs for radiative transfer modeling

Abstract

Radiative transfer models that use spatially explicit 3D models to represent forest structure can simulate highly realistic Earth Observation (EO) data. Such simulations at the forest stand scale (≥ 1-ha) allow for more direct calibration and validation of EO products. Explicitly reconstructing 3D forest structures at scales that can be compared directly with satellite EO data (i.e., dozens to hundreds of meters) is challenging. Reconstructing large forest areas (≥ 1-ha) using a representative subset (i.e., forest subsampling) is a potentially more practical and feasible method. However, the impacts of forest subsampling on radiative transfer (RT) modeling were never formally tested in the spatially explicit forest scene.This study quantified the trade-offs involved in two main subsampling approaches when reconstructing the spatially explicit scene of a real forest for RT modeling. The two subsampling approaches were: (1) subplot subsampling - area-based, using the subplot (i.e. a fixed area) as the basic sampling and reconstruction unit; and (2) tree library subsampling - tree-based, using the individual tree as the basic sampling and reconstruction unit. We used the Discrete Anisotropic Radiative Transfer Model (DART) to simulate the Bidirectional Reflectance Factor (BRF) of the completely reconstructed 1-ha 3D-explicit forest scene, as well as the simplified forest scenes built from various subsets of the same forest. The simulated reflectance deviation of the simplified forest scenes was evaluated by comparing it with the fully reconstructed forest scene.The results showed that for subplot subsampling, as the sampling fraction increased from 10% to 90%, the normalized mean BRF deviation of radiative transfer simulations decreased from −2.7% to −0.0034% and its standard deviation decreased from 7.7% to 0.54%. Additionally, as the sampling fraction increased from 10% to 90%, the normalized mean BRF deviation of tree library subsampling decreased from −7.4% to −1.3% and its standard deviation decreased from 2.8% to 0.51%.Overall, sampling 20% of the forest area using the subplot subsampling method was an effective reconstruction strategy for the temperate deciduous forest. This strategy ensured a normalized mean BRF deviation of −2.9% and a standard deviation of 3.7% in the forest stand studied, with a relatively low reconstruction effort.This study shows that the forest subsampling approach allows for reconstructing 3D-explicit forest scenes for RT modeling at the forest stand scale (≥ 1-ha). Accordingly, highly realistic EO data can be modeled at the forest stand scale, which allows for more direct calibration and validation of EO products.