Dargana: fine-tuning EarthPT for dynamic tree canopy mapping from space

TL;DR

Dargana demonstrates that a pre-trained EarthPT time-series model can be efficiently specialized to perform near-time, 10 m resolution canopy-type mapping (conifer vs broadleaf). By freezing the base model and fine-tuning a lightweight MLP head on NFI data, the approach achieves high accuracy (ROC-AUC 0.98, PR-AUC 0.83) on unseen Cornwall imagery and can reveal fine structures and temporal changes beyond the fine-tuning sample limit. The work showcases scalable, dynamic land-cover monitoring from space with potential for natural capital management, forest management, and conservation. It emphasizes that the main bottleneck is availability of high-quality labeled data, suggesting easy extension to broader regions and land-cover classes as data become available.

Abstract

We present Dargana, a fine-tuned variant of the EarthPT time-series foundation model that achieves specialisation using <3% of its pre-training data volume and 5% of its pre-training compute. Dargana is fine-tuned to generate regularly updated classification of tree canopy cover at 10m resolution, distinguishing conifer and broadleaved tree types. Using Cornwall, UK, as a test case, the model achieves a pixel-level ROC-AUC of 0.98 and a PR-AUC of 0.83 on unseen satellite imagery. Dargana can identify fine structures like hedgerows and coppice below the training sample limit, and can track temporal changes to canopy cover such as new woodland establishment. Our results demonstrate how pre-trained Large Observation Models like EarthPT can be specialised for granular, dynamic land cover monitoring from space, providing a valuable, scalable tool for natural capital management and conservation.

Figures (6)

• Figure 1:
• Figure 2: Dargana comprises of a pre-trained frozen EarthPT base model ($\Phi$) and a trainable MLP head ($\Psi$) that is appropriate for the downstream task at hand. In this study that task is tree canopy type classification.
• Figure 3: Case studies showing (a): evidence of Dargana's ability to track change and (b): information extraction below the spatial limit of the fine-tuning dataset.
• Figure 4:
• Figure 5: Class probabilities for conifer and broadleaved tree types for SX25. (Top) pixels are coloured according to class probability on a logarithmic scaling to emphasise fine structure; (middle) pixels are coloured green where the probability was the maximum across all possible classes; (bottom) National Forest Inventory labels. The right column shows the combination of conifer and broadleaved tree types. Note how Dargana identifies structure not present in the training data. We identify two sub regions 'Trenant' and 'Duloe' that are examined as case-studies in Fig. \ref{['fig_casestudies']}.