Answers are grounded in this paper and cite their sources.
Providing accurate and up-to-date tree species information is critical for estimating, responding to, and anticipating the impacts of climate change, enabling flexible and sustainable forest management strategies. Although large-scale tree species mapping using remote sensing offers advantages in terms of time, cost, and spatial coverage, achieving accuracy comparable with traditional methods such as field surveys remains challenging, particularly when transferring models to large forested territories. While multimodal remote sensing datasets have shown good potential for small-scale tree species classification, automatic approaches using spaceborne hyperspectral imagery (HSI) and airborne laser scanning (ALS) remain unexplored for large-scale boreal forests. This study presents an automatic tree species classification approach using multimodal deep learning (MDL) that combines ALS with PRISMA HSI data in eastern Canada. We developed a MDL approach that integrates a dynamic graph convolutional neural network branch for extracting structural features from ALS data, and a convolutional neural network branch for spectral feature extraction from HSI data. In addition, we evaluated the potential of three PRISMA wavelength regions for tree species mapping. Six common tree species were evaluated: white birch, yellow birch, trembling poplar, black spruce, jack pine, and balsam fir. Results indicated that combining HSI with ALS improved the overall accuracy (OA) by 11.2% compared with ALS alone, reaching an OA of 76.5% across both mixed and pure plots. The achieved accuracy reflects the valuable contribution of the near-infrared and short-wave infrared spectrum region, which provided a 5.6% improvement compared with using only the visible region. The OA reached 84.5% in plots with species purity over 70%. Evaluated species achieved F1-scores between 66.7% and 90.0%. Furthermore, the high spectral resolution of HSI increased OA by 7.7% compared with Landsat data combined with ALS. This OA of 84.5% for plots with species purity above 70% is comparable with conventional photo-interpreted forest inventories in Canada, which typically target accuracies of around 80%–85% for dominant tree species, while our approach additionally provides a classification probability map.
No takes yet. Share an insight, caveat, or question.
Gahrouei et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: