Fusion of European forest inventories with Sentinel-1 and Sentinel-2 data for improving scalability in estimating forest variables

Large scale forest inventory plot data are key to monitor forest ecosystems, but while they provide very detailed information at tree level they are limited in resolution in both space and through time. Earth Observation (EO) data offer the opportunity for scaling up plot data and improving the temporal resolution of monitoring. However, there are significant challenges to this, including small field plot sizes, pre-processing and potential GPS errors in aligning the data, whilst the huge amount and variety of EO data introduce substantial challenges of high dimensionality, in addition to the noise of training and testing data, within any AI system. In this work, we fuse plot and Earth Observation data, demonstrating the value of embedding existing and newly EO derived metrics, and selecting the most important features to improve monitoring of forest properties at large scales. 

In this work we work with Sentinel-1 (SAR) and Sentinel-2 (optical) and inventory data from close to 10,000 plots in Spain, measured from onwards. SAR data require substantial pre-processing due to noise and acquisition, topographic and moisture effects. We used pre-processed SAR data, and filtered for non-shaded slopes, removed plots close to surface water and data collected on days with high precipitation. We masked out clouds from our optical data. After fusing the EO data, we removed disturbed areas using the Global Forest Change Collection and plots with high variability of pixels around them to reduce uncertainty due to the small sizes of the plots. As well as using standard indices (e.g., NDVI, RVI), we derive new metrics of the phenological cycle of the forest from monthly averages of indices and bands by selecting features from peaks and troughs. We reduce dimensionality using principal component analysis and random forest to select the most important features. Chosen features are used for training and evaluating a customized AI system to estimate forest variables such as total basal area, stem density, mean diameter at breast height and forest type. The code implemented in Google Earth Engine JavaScript and Python will be released as open source.