Geospatial computer vision,
research to production.

Data. 1.2 km areas of Puerto Rico at 0.78 m/px, from USGS The National Map orthoimagery, which is public domain. Labels are OpenStreetMap building footprints rasterised to the same grid. San Juan, Caguas, Mayagüez, Carolina and Arecibo were used for training, and Bayamón to choose the prompt and the checkpoint. Ponce and Guaynabo weren’t touched until scoring — they’re what you see above.

Model. SAM 3.1, Meta’s open-vocabulary segmentation model, prompted with “house” — zero-shot on Bayamón it beat “roof” and “building” (0.58, 0.55 and 0.48 IoU). I froze the image backbone and text encoder and trained the detection and mask heads, about 25M of its 840M parameters, for 1,000 steps on the five training towns. The checkpoint kept is the one that did best on Bayamón. Each tile is cut into overlapping 300 m windows so the model isn’t asked to find thousands of buildings at once, and each pixel keeps the highest mask probability × confidence of any detection over it. Every number in the panel is computed in your browser from the model’s actual probability output against the held-out labels — nothing is hard-coded.

What fine-tuning bought. On the same two towns, a compact U-Net I trained from scratch scored 0.50 IoU in Ponce and 0.59 in Guaynabo. SAM 3.1 straight out of the box scored 0.54 and 0.55. Fine-tuned, it’s 0.58 and 0.62 — ahead of both in both towns, with precision and recall each up. It also fixed SAM’s confidence on overhead imagery: zero-shot its best thresholds were 0.15–0.25, and now they’re 0.40–0.45.

What it gets wrong, and why that matters. The labels are not truth, they are OpenStreetMap. Where the map is incomplete, a correct detection is scored as a false positive — so the orange areas are a mix of genuine model error and buildings nobody has mapped yet. Dense blocks of attached roofs and buildings under tree cover are the usual hard cases for footprint models, so that’s where to look. A production system would need higher-resolution imagery, multi-season capture, and human verification on the label set before any of it informed a real decision.

This is an independent demonstration built for this site. It uses no client data, no project code, and no federal deliverable.