Data assimilation of forest status using Sentinel-2 data and a process-based model

Abstract

Spatially explicit information of forest status is important for obtaining more accurate predictions of C balance. Spatially explicit predictions of forest characteristics at high resolution can be obtained by Earth Observations (EO), but the accuracy of satellite-based predictions may vary significantly. Modern computational techniques, such as data assimilation (DA), allow us to improve the accuracy of predictions considering measurement uncertainties. The main objective of this work was to develop two DA frameworks that combine repeated satellite measurements (Sentinel-2) and process-based forest model predictions. For the study three tiles of 100 × 100 km² were considered, in boreal forests. One framework was used to predict forest structural variables and tree species, while the other was used to quantify the site fertility class. The reliability of the frameworks was tested using field measurements. By means of DA we combined model and satellite-based predictions improving the reliability and robustness of forest monitoring. The DA frameworks reduced the uncertainty associated with forest structural variables and mitigated the effects of biased Earth Observation predictions when errors occurred. For one tile, Sentinel-2 prediction for 2019 (s2019) of stem diameter (D) and height (H) was biased, but the errors were reduced by the DA estimation (DA2019). The root mean squared errors were reduced from 5.8 cm (s2019) to 4.5 cm (DA2019) and from 5.1 m (s2019) to 3.3 m (DA2019) for D (sd = 4.33 cm) and H (sd = 3.43 m), respectively. For the site fertility class estimation DA was less effective, because forest growth rate is low in boreal environments; long term analysis might be more informative. We showed here the potential of the DA framework implemented using medium resolution remote sensing data and a process-based forest model. Further testing of the frameworks using more RS-data acquisitions is desirable and the DA process would benefit if the error of satellite-based predictions were reduced.