Prithvi-EO-2.0 Crop Classification Performance Evaluation Across Continents: Unveiling Model Reliability Challenges Unde

Summary:This study evaluates the out-of-distribution performance of the widely adopted geospatial foundation model Prithvi-EO-2.0 across 37 events in 12 countries spanning three continents. The results show a decline in mean overall accuracy (OA) from 0.65 in the United States to 0.40 in Europe, highlighting significant geographic transferability issues. The research further assesses the model's robustness to observation window misalignment, class scheme coarsening, band loss, and cloud and shadow conta

---

Background and Motivation

Geospatial foundation models (GeoFMs) have demonstrated strong capabilities in crop classification using large-scale satellite data sources, but their cross-regional transferability and operational reliability require deeper evaluation. This study aims to systematically assess the performance of Prithvi-EO-2.0 across continents and provide recommendations to enhance its deployment reliability.

Key Findings and Results

1. Geographic Transferability Issues: The mean overall accuracy (OA) of Prithvi-EO-2.0 in the United States is 0.65, while it drops to 0.40 in Europe, indicating significant geographic transferability challenges.
2. Phenological Influence: Accuracy collapses when the observation window misaligns with local crop phenology, while deterministic confidence remains high.
3. Robustness Assessment:
   • Class Scheme Coarsening: Consolidating 13 classes into 10 increases the mean OA by 8.4 percentage points.
   • Observation Window Compression: Compressing the window to 45-90 days, peaking near 75 days, preserves accuracy, while windows tighter than 30 days fall about 0.11 below that plateau.
   • Other Factors: Band loss and cloud and shadow contamination also affect model performance, but to a lesser extent than phenological misalignment.

Recommendations for Improvement

1. Adjust Observation Windows: Align observation windows with local crop growing seasons to minimize the impact of phenological misalignment on model performance.
2. Optimize Class Schemes: Adjust class schemes based on the model's dominant confusions to improve classification accuracy.
3. Enhance Robustness: Incorporate more diverse data sources and more complex scenarios in model training to improve the model's adaptability to different geographic and climatic conditions.

Industry Impact and Future Directions

This study provides critical insights for the practical deployment of geospatial foundation models, emphasizing the importance of model adaptability across different geographic and phenological conditions. Future research could further explore how multimodal data fusion and adaptive learning mechanisms can enhance model performance in complex environments.

---

Source: ArXiv Machine Learning (cs.LG) (2026-10-08)

— END —

Tags: #Geospatial Models Follow #Crop Classification Follow #Model Evaluation Follow #Cross-Regional Transferability Follow #Prithvi Follow