Speaker
Description
This work analyses the possibility of estimating forest tree height (TH) at high and low spatial resolution using Machine Learning (ML) models and multi-source Earth Observation (EO) data. Estimating forest biophysical parameters is a must to quantify climate change effects, and the challenge with EO is the complexity and volume of data. Building on the observation that a direct relation exists between the SAR interferometric coherence, especially the temporal decorrelation component, and the biophysical parameters of forests, this work studies the possibility and the effects of combining interferometric coherence with artificial intelligence models, AI4EO, to estimate forest tree height. The interferometric coherence along a digital elevation model (DEM) and a land cover map (LC) is combined with meteorological data in a Random Forest model to estimate tree height in three different test sites: the Pyrenees region between southern France and Catalonia in Spain, the central part of Finland and an area close to Yellowknife, Canada, in the Northwest Territories. The three areas expand across 200km × 50km, and for each of them, 20 Sentinel-1A and Sentinel-1B acquisitions separated 6 days apart are used, obtained from December to March for different years depending on the site, with only VV polarization processed for the 6-day complex interferometric coherence computation. The ground-reference data consist of a DEM from Copernicus 30m Global DEM, an LC map and a tree height ground reference, at a high spatial resolution of 10m–30m. The meteorological data are obtained through ERA5 and include rain, snow, wind and temperature at a lower spatial resolution of 10km as a consequence of the model; in total, 8 different variables are used. The study consists of two parts. First, the estimation of tree height at the low 10km×10km resolution by downscaling the high-resolution variables, aggregating the pixels at 10km × 10km subareas, assessing the contribution that the InSAR coherence at 10km may bring when combined with the native ERA5 meteorological data, DEM and LC. Second, a high-resolution study at 10m × 10m, to study if the weather conditions add complementary information to the native Sentinel-1 resolution data; to reduce computational complexity, this analysis is limited to a 10km × 10km subset and only a single 6-day coherence pair is employed. The model employed is Random Forest, as it outperformed other tested models, such as Support Vector Machines and K-Nearest Neighbours. In the low-resolution analysis the model is trained and tested using each individual 6-day coherence acquisition, and the mean of the results defines the validation metrics: the mean Root-Mean-Square-Error (RMSE) is 2.098m, 1.231m and 2.009m and the mean determination coefficient (R2) is 0.702, 0.211 and 0.690 for the Pyrenees, Finland and Yellowknife, respectively. The 10m analysis uses just the first 6-day coherence acquisition of each site; hence no mean is required; the RMSE is 4.442m, 3.053m and 1.698m and the R2 is 0.672, 0.549 and 0.837, respectively. Relatively low RMSE and R2 are achieved for both analyses, although the errors for the Pyrenees and Finnish sites double when moving from the low-resolution to the high-resolution analysis, primarily due to the homogeneity introduced in the low-resolution data, while for the Canadian site the error decreases by approximately 20 %. These errors must be interpreted in the context of their spatial resolutions, thus in the low resolution, the variability in heights is eliminated. The results show how the combined information from these multiple sources complements each other to estimate the tree height with a relatively low error in an operational way, highlighting the potential of combining multi-source data and machine learning techniques and opening the possibility of further data fusion for improved results.