National Tree Species Mapping in Denmark with Machine Learning Using Multi-Temporal Sentinel Data

Forests play a critical role in sustaining biodiversity, regulating carbon and water cycles, and providing aesthetic and amenity value. While much is already known about the distribution of forest cover, detailed tree species composition remains poorly mapped at national scales. Earth observation, particularly through Sentinel missions, provides dense multi-temporal and multi-sensor data that, when combined with machine learning, can enable improved characterization of forest composition [1]. This study presents a machine learning assessment of tree species mapping across Denmark using multi-temporal Sentinel-1 and Sentinel-2 data. 

The task is formulated as a pixel-based, time-series multi-class classification problem. Two input representation strategies are evaluated: i) manually engineered features incorporating spectral bands, vegetation and moisture indices processed from Sentinel data, ii) pre-computed embeddings from Earth Observation foundation models (AlphaEarth [2] and Tessera [3]), which encode spatio-temporal information from multi-source Earth observation data. Both input representations were complemented by canopy height information from national elevation data provided by the Danish Agency for Data Supply and Infrastructure. Random Forest, XGBoost, and Artificial Neural Network classifiers were trained and evaluated for each representation using species-level reference data from the Danish National Forest Inventory. 

Results show that the traditional feature engineering approach achieves strong performance for tree species mapping, with consistent gains from Sentinel-1/2 fusion. Foundation model embeddings yield comparable, though slightly lower, accuracy under full training data conditions. However, in data-limited training scenarios, they outperform the feature-based workflow, indicating increased robustness to reduced training sample sizes. Moreover, the use of pre-computed embeddings reduces processing complexity and computational requirements by removing the need for data preprocessing and manual feature engineering, yielding benefits that extend beyond performance alone. 

Our findings highlight the effectiveness of machine learning for national-scale tree species mapping using Sentinel data and provide new evidence that Earth observation foundation model representations offer viable alternatives to handcrafted features. The study contributes to advancing operational forest monitoring and provides insights into the integration of foundation models into large-scale ecological mapping workflows. 

References 

[1] Holzwarth, Stefanie, Frank Thonfeld, Patrick Kacic, Sahra Abdullahi, Sarah Asam, Kjirsten Coleman, Christina Eisfelder, Ursula Gessner, Juliane Huth, Tanja Kraus, and et al. 2023. "Earth-Observation-Based Monitoring of Forests in Germany—Recent Progress and Research Frontiers: A Review" Remote Sensing 15, no. 17: 4234. https://doi.org/10.3390/rs15174234 

[2] Brown, Christopher F., Michal R. Kazmierski, Valerie J. Pasquarella, et al. “AlphaEarth Foundations: An Embedding Field Model for Accurate and Efficient Global Mapping from Sparse Label Data.” arXiv:2507.22291. Preprint, arXiv, September 8, 2025. https://doi.org/10.48550/arXiv.2507.22291. 

[3] Feng, Zhengpeng, Clement Atzberger, Sadiq Jaffer, et al. “TESSERA: Precomputed FAIR Global Pixel Embeddings for Earth Representation and Analysis.” arXiv:2506.20380. Preprint, arXiv, September 22, 2025. https://doi.org/10.48550/arXiv.2506.20380.