SEES 2026 Land Cover Archive: Lower Manhattan, New York City

Geospatial transformer satellite models are crucial for Earth observation, disaster response, and zoning policy. However, satellite data can be noisy in densely urban areas such as Manhattan, New York City. Key issues such as building occlusion can affect visible foliage levels, decreasing model classification ability. Using the GLOBE Observer program, we collected 3,600 ground samples (5-meter footprints) across 36 Lower Manhattan sites and paired them with Harmonized Landsat-Sentinel imagery. Using this data, we fine-tuned the Prithvi-EO-2.0 Foundation model to improve urban land cover mapping. We examine model performance change across two labeling strategies: hard labels, which force each mixed pixel into one class, and soft (fractional) labels, which train the model toward the true built-up fraction. To combat severe class imbalance towards built-up labels, we evaluate class-weighted and focal-loss training and report precision-recall performance for the minority class. Our fine-tuning method improved separability over the pretrained representation. Our results suggest that volunteer-collected labels, paired with foundation models and sub-pixel-aware training, can extend fine-grained urban land-cover mapping to data-sparse cities. This represents a working draft of our JEOGA paper that will be completed in the future. This work was completed throughout June, July, and August 2026.