Deforestation Attribution using the Prithvi E.O. Foundation Model

Attributing deforestation to key natural and anthropogenic drivers is critical for aiding reliable, transparent, and effective forest loss mitigation and management strategies. In recent years, the rapid evolution of Geospatial Foundation Models (GFMs) provides a unique opportunity for a fast, large-scale, and computationally efficient detection and attribution of forest loss. In this study, we evaluated the capability of the Prithvi-EO-2.0 300M and 600M GFMs for attributing primary deforestation events to four major drivers – grassland-shrubland expansion, plantation, smallholder agriculture and others (mining, fishpond, logging roads, secondary forests, etc.). We adapted the ForestNet dataset using a Stratified Spatial Shuffling for the training, validation, and test splits and a Scene Data Augmentation strategy for developing a geographically robust and well-balanced dataset. Using the Prithvi-EO-2.0 models, we compared three finetuning strategies – linear probing, low rank adaptation (LoRA, with ~2% of trainable parameters), and full fine-tuning with different fractional data inputs against a baseline ResNet Feature Pyramid Network trained end-to-end with 1.) imagery inputs only; and 2.) imagery and auxiliary inputs. Across all data fractions, Prithvi-EO-2.0 with low rank adaptation and the full-finetuning adaptation strategies outperforms the baseline model, however the linear probing underperforms in certain cases (by macro-averaged accuracy). The best performing models for the imagery only set up were the Prithvi-EO-2.0 600M model using a LoRA strategy (80.40%), followed by the full finetuning strategy (80.22%) for the same model, marginally surpassing the baseline (80.21%). However, when trained with additional auxiliary datasets, the Prithvi-EO-2.0 300M and 600M adapted with a full finetuning strategy reached an 82.75% and 82.94% accuracy, with a +3.8-percentage point accuracy gain over the baseline. The LoRA strategy maintains ≥75 % accuracy at full data for both the 300M and 600M models, while updating just 11.3 million trainable parameters, demonstrating its low‑resource efficiency. These results show the potential of the Prithvi EO models to deliver accurate, computationally efficient deforestation attribution, useful for large-scale forest change attribution.