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

Abstract

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