Justin Guthrie

A global, open-source benchmark for classifying critical infrastructure from space: 18,756 tiles spanning seven continents, built on Sentinel-1 and Sentinel-2 imagery with OpenStreetMap-derived weak labels, evaluating seven foundation model backbones across 10 infrastructure classes. The strongest backbone reaches 57.9% macro F1 against a 39.2% ResNet-18 baseline. The more interesting result is the spread across sectors: transport and data centers are read reliably from orbit (airports 85.3%, train stations 82.1%, data centers 77.6%), while power-sector classes sit between 27.5% and 46.2%. Which infrastructure is legible from orbit turns out to depend sharply on what kind of infrastructure it is.

Deep learning detection of electricity transmission towers at national scale, with Ed Oughton at George Mason University. More to come.

Investigating interpretability of Google DeepMind's 64-dimensional AlphaEarth satellite embeddings, identifying which dimensions are most predictive for ESA WorldCover land cover classification. The embedding space turns out to be functionally organized but substantially redundant: land cover performance concentrates in a small subset of the 64 dimensions, and which dimensions carry that signal varies by class. Built a full-stack interactive dashboard (React/TypeScript, Python/Polars) for exploring 130K+ experimental records.

Cited in Camps-Valls et al. (Nature Communications, 2026) as one of two examples of mechanistic interpretability revealing internal representations in Earth-system foundation models, alongside work on GraphCast; and in Stewart et al., "Earth Embeddings" (arXiv:2608.03410, 2026).

Updated three chapters from their 2017 publication: 'Climate Extremes and Informing Adaptation,' 'Informing Climate Adaptation with Earth System Models and Big Data,' and 'Climate Adaptation, Introduction.'

Animated simulations in ArcGIS Online demonstrating how hurricane scenarios propagate through coupled power and water networks. Identified single-point-of-failure chokepoints where one circuit loss cascades into water supply disruption, school power outage, and residential service loss. Completed NSF I-Corps Spark Program; conducted customer discovery with Maine municipalities.

Built a Python and Google Earth Engine pipeline over VIIRS Black Marble nighttime lights (VNP46A2) and EAGLE-I outage data to test whether nighttime lights can independently verify California Public Safety Power Shutoff de-energization events, using CPUC post-event reports as ground truth. Validated on Hurricane Beryl in Houston, then applied to the SDG&E December 2024 PSPS event. County-level correlation is strong for the hurricane (r ≈ −0.76) and collapses for PSPS events (r ≈ 0.02). That gap is a resolution mismatch rather than a dead end: shutoffs de-energize small rural pockets inside large counties whose bright urban cores dominate the county-level signal, which is precisely what motivates moving to circuit-level SCE geometry.

Part of the NASA Disasters Program PRISM-grid effort. Coauthor on an abstract submitted to AGU Fall Meeting 2026, "Science for a world of socio-ecological-technological systems: Quantifying and qualifying multi-hazard risk for power grid resilience" (submitted, not yet accepted).

Blending quantitative and qualitative analysis into StoryMaps that communicate infrastructure and economic development findings in a form local and state policymakers can act on, across 5+ state engagements, with the furthest-along engagement presented to a state secretary of commerce. Earlier work developed a geodatabase of 300+ U.S. shipyards and repair facilities with 10+ analytical attributes per site, built an interactive ArcGIS Dashboard for internal stakeholder use, managed geospatial data sharing across 20+ users, and cleaned and visualized social capital investment data across U.S. metropolitan areas.

Enhanced real-time Python data pipeline in Azure for track change handling across the MBTA Commuter Rail, cross-referencing Amtrak and MBTA schedules. Implemented NetworkX graph model using Dijkstra's algorithm; established framework for South Coast Rail integration. Improved scheduling accuracy for 81K+ daily passengers via XML transformations across 135 stations.

Collected baseline resilience indicators (transportation, public safety, urbanism) for three communities in St. Thomas, USVI using the I-RES methodology. Created animated population migration timeline from U.S. Census data. Developed StoryMaps for Caribbean infrastructure challenges. Assisted planning collaborative sessions integrating local stakeholder input into resilience assessments.

Created 3D panorama tour of Porto Novo, Benin for the African Built Heritage Hub using 3DVista. Processed LiDAR point cloud data into mesh layers for urban-scale 3D models in ArcGIS Online.