|Paper title||A forest digital twin underpinning the validation of Sentinel-2 fAPAR|
|Form of presentation||Poster|
Digital twins are becoming an important tool in the validation process of satellite products. Many downstream satellite products are created based on a complex chain of processing procedures and modelling techniques. Vegetation biophysical products are a classic example of this, particularly in forests where the 3D arrangement of canopy constituents is heterogeneous and its variability across different forest types is high. This means that satellite product algorithms applied to forests employ a range of assumptions about the forest constituents and illumination characteristics in order to best estimate quantities such as the fraction of absorbed photosynthetically active radiation (fAPAR) and leaf area index (LAI). This leads to a definition difference between the quantity being (assumed to be) measured by the satellite sensor and that which is actually measured on the ground using in situ measurement techniques (which might also have their own assumptions). Simulation studies using digital twins offer a way to overcome these issues.
This contribution describes an fAPAR validation exercise of the Sentinel-2 fAPAR product over Wytham Woods (UK) for 2018. It combines in situ measurements of fAPAR with correction factors derived from radiative transfer (RT) simulations on a digital twin of Wytham Woods. The digital twin (which is open source) is based on datasets collected during the summer and winter of 2015/2016 and represents a 1 ha area of temperate deciduous forest. The leaves and stems are derived from LiDAR point clouds collected every 20 metres throughout the forest and combined with spectral measurements of the respective canopy and understory components (bark, leaves, soil, etc.). This model represents a useful surrogate with which to test canopy configurations and forest structure assumptions that are impossible at the real study site. As an example, in certain satellite fAPAR products it is assumed that only photosynthesising elements are present in the canopy (green fAPAR). To analyse a situation such as this, in the model we can remove the stems and branches from the RT simulations and compare that to simulations on the full model to assess the differences.
Combined with this, we use a PAR network located at Wytham Woods to derive fAPAR. Each sensor in this network is calibrated and produces results that have a well characterised uncertainty and are traceable to SI. Using the Wytham Woods digital twin we are able to simulate a reference fAPAR value would be under a specific set of illumination conditions since it is possible to track the fate of each photon/ray in the scene. As a result, we have a form of traceability defined as virtual traceability to this reference.
Using the measurement and modelling component discussed above, we were able derive correction factors for the satellite and in situ measurements (relative to the reference value). Allowing the in situ and satellite values to be compared through a common intermediary. The results show that the correction factors reduced the deviation between the in situ and satellite-derived fAPAR. Since the digital twin is representative of the summer months (leaf-on), the deviations (post-correction) are largest in the winter, with a quick decrease in the spring (with leaf production) and a slow increase from July to October as senescence takes place.
This work provides a highly detailed look at a single forest location and single satellite product. Given the large biases found, and corrected for, it suggests that future work is required to understand how these biases (and subsequently the correction factors) change in space (e.g. for different biomes, etc.), time and for different satellite products. This means that the fAPAR (and other vegetation related satellite products) community should create many more forest digital twins to facilitate this. This is a top priority if we are to reach the GCOS requirements for fAPAR (measurement uncertainty of < (0.05 or 10%)) and, more importantly, if downstream users of these products are to trust them.