Use of Sentinel-2 Data to Improve Multivariate Tree Species Composition in a Forest Resource Inventory

Aerial-photo interpreted inventories of forest resources, including tree species composition, are valuable in resource management, but expensive to create and can be relatively inaccurate. Because differences among their spectral properties seasonal phenologies, it might possible improve such inventory information (FRI) by using concert with multispectral satellite from multiple time periods. We used Sentinel-2 nine bands 12 dates within a two-year period model multivariate percent composition >51,000 stands the FRI south-central Ontario, Canada. Accuracy random (RF) convolutional neural network (CNN) predictions were tested species-specific basal area 155 0.25-ha field plots. Additionally, we created models data compared accuracy these FRI-based use areas second (13.7-ha) set. Based on average R2 values across two sets, Sentinel-FRI outperformed FRI, showing 1.5- 1.7-fold improvements relative for RF 2.1- 2.2-fold CNN (mean R2: 0.141–0.169 (FRI); 0.217–0.295 (RF); 0.307–0.352 (CNN)). Models performed even better: 2.1-fold 2.8-fold 0.169 0.356 0.469 As predicted, between FRI- field-trained higher than FRI. Of 21 evaluated, 8 rare had poor all cases. Our approach allowed us more creation if been restricted dominated single map abundances at resolution. It further stem maps incorporation effects canopy disturbances.