Geospatial Foundation Models
AI-Powered Earth Observation and Satellite Imagery Analysis
Key Results
93.75%
Overall accuracy
Random forest on held-out 2023 patches from frozen embeddings with no fine-tuning (90.8% balanced accuracy, macro-F1 0.898). Logistic regression, random forest, LightGBM and XGBoost all land within 0.27 percentage points of each other — the signal is in the embeddings, not the classifier.
95.3%
Agreement with human reference
Against the 385-point blind human reference, versus 91.7% for the USDA CDL labels the model was trained on (McNemar exact p = 0.016). The model agrees with expert interpreters more often than its own training labels do.
92.5–94.3%
Cross-year transfer (2018→2023)
All 36 train/evaluate year pairs stay in this band with no retraining.
140× fewer labels
Label efficiency
60,000 balanced pixels instead of the ~8.6 million pixel pool costs 1.3 pp of overall accuracy and actually gains 1.8 pp of balanced accuracy.
95.3% vs 93.5%
vs fine-tuned TerraMind
Competitive with a fine-tuned TerraMind model on the same points (p = 0.14, i.e. not a significant difference) at a fraction of the training cost. TerraMind predictions come from Giovanni Montefoschi's parallel thesis on the same dataset.
Defended July 2026
Thesis outcome
144-page thesis approved 2 July 2026, defended 22 July 2026.
Thesis at a Glance
Key Insights
- 93.75% accuracy from frozen embeddings — no fine-tuning needed
- Agrees with expert interpreters more often than its own training labels (95.3% vs 91.7%)
- Transfers across years 2018–2023 with no retraining (92.5–94.3%)
- 140× fewer labels at a cost of 1.3 points of overall accuracy
Research Timeline
Literature Review Started
June 2025
TerraMind & AlphaEarth Studies
September 2025
Experiments & Blind Human Validation
2026
Thesis Approved
2 July 2026
Thesis Defended
22 July 2026
arXiv Preprint & SpringerBriefs Chapter
Planned
Supervision & Publications
Supervision
Prof. Vasil Yordanov
Supervisor
Politecnico di Milano
Dr. Zhongxin Chen
Co-supervisor
FAO-UN
Publications
Mohammad Ammar Mughees, "Binary Cropland Classification from AlphaEarth Embeddings: A Geospatial Foundation Model Approach for Maine, USA", MSc thesis, Politecnico di Milano, 2026. 144 pp.
Approved 2 July 2026 · defended 22 July 2026arXiv preprint of the thesis work.
Planned"From Foundation Embeddings to Cropland Maps: Label Efficiency, Transferability, and Human Validation" — chapter for the Politecnico di Milano SpringerBriefs series.
Planned
Future Work
TESSERA Comparison
TESSERA, from Cambridge researchers, provides precomputed FAIR global pixel embeddings for Earth representation and analysis — a natural comparison point for AlphaEarth. Comparison notebooks (12–15) were built during the thesis but deliberately kept outside the submitted scope, making this the first candidate for follow-up work.
Research Potential:
- • Comparative analysis with AlphaEarth and TerraMind
- • Architecture differences and performance benchmarks
- • FAIR (Findable, Accessible, Interoperable, Reusable) approach impact
- • Real-world application scenarios and limitations
TerraMind GFM
Comprehensive Analysis
Deep dive into TerraMind's geospatial foundation model architecture, capabilities, and applications in satellite imagery analysis.
AlphaEarth
Google DeepMind Embeddings
Analysis of Google's AlphaEarth foundation model for global mapping from sparse label data and its breakthrough innovations.
Methodology
Research Approach
Detailed research methodology, comparative analysis framework, and evaluation metrics for geospatial foundation models.