I turn satellite and drone imagery into measurements people can check.

Land cover with honest accuracy

Random-forest land-cover classification of Bryan–College Station from Sentinel-2, showing how random cross-validation overstates accuracy compared with spatial block validation, with area estimates and 95% intervals (Olofsson et al. 2014).

Held-out agreement with ESA WorldCover: 0.81 ± 0.015 (area-weighted, 95% CI). With polygon-like training patches, random cross-validation scored 0.89 balanced accuracy but held-out blocks only 0.83; all six area estimates' 95% intervals contained the reference area.

Open Earth MCP server

A Model Context Protocol server that lets Claude or any MCP client look up open Earth-observation data (Sentinel-2 NDVI time series, land cover, elevation) and returns the source scenes and dates with every answer, plus a verify check.

Six tools, one resource and one prompt; 29 offline tests plus a live test. Its verify tool recomputes NDVI from the raw bands by a second code path: 0.1685 vs 0.1683 on a Brazos River cropland field whose 2026 season it traced from 0.18 in March to a 0.91 peak in June.

Houston heat hot spots

Where is Harris County hottest in summer, and how much does vegetation explain? Landsat land surface temperature by census tract, Getis-Ord Gi* and LISA clusters, and a spatial regression that accounts for neighbors.

Summer surface temperature is strongly clustered across 1,111 tracts (Moran's I 0.62). After accounting for neighbors, each +0.1 in tract NDVI is associated with about 1.9 °C lower surface temperature (an association, not a causal effect); 147 tracts are significant hot spots after false-discovery correction.

Accuracy Assessment Toolbox for ArcGIS Pro

A Python toolbox that adds what Compute Confusion Matrix leaves out: stratified sample design, area-adjusted accuracy with 95% confidence intervals (Olofsson et al. 2014, with an optional Wilson interval), and buffered spatial blocks for train/test splits. Tested inside ArcGIS Pro 3.7.

Reproduces the class areas in Olofsson et al.'s worked example to the hectare. In a synthetic demo, pixel counting overstated cropland by 39%; over 1,000 repeat samples the 95% intervals covered the true area 95% of the time for four of five classes, and the report flags the one that fell short.

Find places that look alike

Query-by-example search over Sentinel-2 imagery with a self-supervised foundation model (SSL4EO-S12): pick one place and retrieve the places that look most like it, then measure how often the matches share its land cover against a spectral baseline.

Over 7,974 chips of the Brazos Valley, the foundation model and a simple spectral baseline scored about the same on coarse land-cover classes (class-balanced precision@10 0.85 vs 0.83); the model's edge was on rare classes such as water (0.87 vs 0.78).

COVID-19 in Texas: where hot spots emerged

Weekly case and death rates for all 254 Texas counties, March 2020 to February 2023, smoothed for small populations (Empirical Bayes), with Getis-Ord Gi* by wave, a space-time hot spot analysis with Mann-Kendall trends, and an interactive map with a monthly time slider.

174 of 253 analyzed counties were a significant hot spot in at least one month, peaking at 76 in December 2020; in the Omicron waves hot spots returned to South Texas. One county whose reported cases exceeded its population (likely non-resident workers) is excluded, with a sensitivity run that keeps it. Nonmetro counties had a higher death rate than metro counties in every wave (538 vs 294 per 100,000 overall, not age-adjusted).

Segment Anything on aerial imagery

Extracting buildings and fields from 0.6 m NAIP imagery with Meta's Segment Anything Model, then scoring the outlines against reference building footprints at the object level (IoU ≥ 0.5) against a simple baseline.

On a held-out area, text-prompted SAM (Grounding DINO + SAM) reached building F1 0.69 (95% CI 0.62–0.74) against Microsoft footprints, vs 0.50 (0.42–0.57) for an NDVI/brightness baseline; it was precise (0.85) but missed about 40% of buildings.

How fast can a fire engine reach you?

Drive-time service areas from every fire station in Brazos County, Texas, over the OpenStreetMap road network and 2020 Census blocks, then a maximal covering location model (the open-source counterpart of Network Analyst's location-allocation) that picks where one or two new stations would cover the most residents.

82.3% of residents are within a 4-minute drive of one of 25 stations and 96.5% within 8. OpenStreetMap listed only 16 stations; adding the 9 missing ones from official department lists changed the answer, since the OSM-only "best new sites" landed about 200 m from real stations. Solver results match an exhaustive search.

Critical facilities in Texas flood zones

A spatial-SQL pipeline (DuckDB, with PostGIS and SQL Server equivalents) that loads FEMA flood hazard maps and hospitals, nursing homes, schools and fire/EMS stations for five Gulf Coast counties, cleans them, and screens which sit in or near mapped flood zones, cross-checked against GeoPandas on every run.

250 of 2,860 facilities (8.7%) sit in FEMA's 1%-annual-chance flood zone, including 13 hospitals with 2,367 beds and 132 schools with about 90,000 students; Galveston County has the highest share (20%). Every count and distance matches an independent GeoPandas recomputation.

Agentic site selection: where to open a coffee shop

A manager agent and five subagents answer plain-English site questions for Harris County by calling ArcPy tools through a local MCP server, comparing an expert-weighted overlay with Esri's USA Geodemographic Embeddings (Find Similar Features Using Embeddings), and testing both on shops hidden from them. Inspired by the GeoAI demos at the 2026 Esri Energy Resources GIS Conference.

On spatially blocked hold-outs the embeddings reached AUC 0.74 and the overlay 0.71 (random 0.50), but the two methods' top 10% of hexes barely overlap (Jaccard 0.02), and in the urban core the embeddings fell to near random. A numpy re-implementation reproduces Esri's tool exactly (Spearman 1.000).

Daylight in College Station

When does it get dark across the academic year? One self-contained page solves the sun's position in the browser with the NOAA solar-position equations, checked against U.S. Naval Observatory tables (within about 20 s).