Speaker
Description
Accurate tree species identification in young boreal forests is critical for sustainable forest management, biodiversity monitoring, and bioenergy resource assessment. However, current remote sensing approaches predominantly rely on single-sensor methodologies, limiting classification accuracy in complex young stands where trees exhibit similar heights, small crowns, and subtle spectral differences.
This research aims to develop and validate ForestFusionNet, a scalable multimodal deep-learning framework that fuses LiDAR-derived 3D structure, high-resolution digital aerial photogrammetry (DAP), Sentinel-2 multispectral time series, and Planet high-resolution multi-temporal imagery to classify tree species at the individual-tree level in young boreal forests. The work pursues two main objectives: (1) to build a harmonized multi-sensor dataset and extract complementary features, and (2) to design a unified fusion deep-learning architecture and rigorously evaluate its accuracy and computational efficiency.
The study will be conducted in two young boreal forest sites in Juva (27.865E, 61.730N) and Heinävesi (28.736E, 62.509N), Finland. The proposed ForestFusionNet architecture will employ sensor-specific encoders: a point-cloud network for LiDAR and a CNN for aerial imagery.
The proposed ForestFusionNet architecture will utilize sensor-specific encoders, including a point-cloud network for LiDAR, a CNN for aerial imagery, and a temporal module for satellite time series. Different fusion mechanisms (late, intermediate, and deep) will be explored to combine modalities optimally. The model will be trained and validated using field-collected individual-tree reference data. The model's performance will be compared with single-sensor results by evaluating how well it identifies each species using metrics such as precision, recall, F1-score, and overall accuracy. Ablation studies will quantify the contribution of each modality, and computational costs (training/ inference time, memory usage) will be reported to assess deployability for large-area mapping.
We anticipate that ForestFusionNet will achieve significantly higher classification accuracy compared to any single-sensor approach by leveraging the complementary strengths of 3D structure, high-resolution texture, and phenological spectral patterns. The project will deliver (i) an open-access harmonized multi-sensor dataset hosted on Zenodo; (ii) the complete ForestFusionNet code released on GitHub; (iii) empirical evidence on the accuracy-efficiency trade-offs of multi-sensor fusion; and (iv) a scalable operational framework for precision inventory of young boreal forests.
The results will directly benefit forest owners, small- and medium-sized enterprises, and policymakers by enabling cost-efficient, high-precision forest inventory. Moreover, improved individual-tree species information will support more reliable biomass and carbon stock estimates, thereby enhancing national and EU climate reporting under the LULUCF sector and supporting the targets of the European Green Deal.
Keywords: Deep learning, multimodal fusion, tree species classification, young boreal forests, remote sensing
| Keywords | Deeplearning; fusion; classification; forestry |
|---|
