Mapping tree species proportions from satellite imagery using spectral–spatial deep learning

Remote sensing can be used to collect information related forest management. Previous studies demonstrated the potential of using multispectral satellite imagery for classifying tree species. However, methods that map species in mixed stands on a large scale are lacking. We propose an innovative method mapping proportions Sentinel-2 imagery. A convolutional neural network was quantify per-pixel basal area considering neighbouring environment (spectral–spatial deep learning). nested U-shaped (UNet++) architecture implemented. produced entire Wallonia Region (southern Belgium). Nine or groups were considered: Spruce genus, Oak Beech, Douglas fir, Pine Poplar Larch Birch and remaining The training dataset model prepared parcels extracted from public administration’s geodatabase Wallonia. accuracy predicted covering region independently assessed data regional inventory robust assessment maps proposed assessing (1) majority species, (2) composition (presence absence), (3) (proportion values). achieved value indicator OAmaj (0.73) shows our approach pure stands. Indicators MS (0.89), MPS (0.72) MUS (0.83) support predict most cases study area. fir best results, with PAs UAs close higher than 0.70. Particularly, high performance detecting genus Beech low proportions: 0.70 0.4 proportion. Predicted had Radj2 0.50. method, which uses spectral–spatial learning is because it adapted complexity forests spatial resolution current Additionally, optimises use available conception by all pixels highly When inventories broad sense, is, georeferenced areas this reproducible applicable at scale, offering