Roof Intelligence Methodology

First, check the source.

Each normalized source record carries a source name, date when known, warnings and a mode: live, demo or unavailable. An analysis timestamp is not an imagery acquisition date. Missing observations stay unavailable; a model assumption does not become a measurement because a calculation uses it.

Demo Data means synthetic/sample properties and scenario inputs. It is not a survey of actual Tempe roofs, a customer list or evidence of savings. Live mode must not silently substitute demo observations when credentials, coverage or an upstream service fail.

Tempe is the initial target market, not a claim of a completed customer pilot. A radius search is not a municipal boundary. Boundary and building-source provenance must be checked independently.

Datasets and their limits

Earth Engine: Server-side geospatial processing requires a registered Cloud project, appropriate permissions and the applicable commercial or noncommercial access configuration. Its availability does not establish the availability or suitability of every dataset.[4] Heat, vegetation and solar context should retain the selected collection, time window and processing scale. Regional land-surface temperature is not indoor temperature or a direct roof thermometer.
AlphaEarth Foundations: GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL represents annual surface conditions with 64-dimensional embeddings on a 10-meter grid. Bands must be interpreted together rather than as standalone physical quantities.[1] A pixel may mix roof, road and vegetation. Use it for learned geospatial similarity or future validated classifiers—not to identify membrane chemistry, roof age or reflectance. The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind.
Google Solar API: Building Insights provides building dimensions, roof segments, solar potential and imagery metadata where covered. The returned building can differ from a geocoded location.[3] Check the match, area, segment consistency, imagery date and quality; preserve missing coverage and ambiguous selections. Photovoltaic production and panel-financing outputs are not inputs labeled as cool-roof savings.
USDA NAIP: USDA/NAIP/DOQQ offers aerial imagery with acquisition-dependent resolution and age. The catalog describes one-meter acquisitions, some older two-meter imagery, and lists 0.6-meter bands; inspect the actual asset rather than promising one resolution everywhere.[2] Credit USDA Farm Production and Conservation – Business Center, Geospatial Enterprise Operations. Imagery access, display, analysis and redistribution require separate rights checks.

Property selection and visual observations

An address search or map click supplies candidate coordinates. A nearby address is not proof of a building match; confirm the highlighted geometry and resolve multiple structures before relying on roof area. Geometry, environmental context and imagery can have different dates and spatial support.

When permitted imagery is available, Gemini can provide an AI-assisted visual observation with field-level confidence, image date and warnings. “Unknown” is a valid result. Apparent brightness, shade and obstructions are visual proxies—not certified reflectance, exact material, structural condition, age or warranty status. Confidence is not a validated probability of correctness.

Google basemap screenshots are not an automatic source for AI processing. A display license does not establish permission for machine analysis or report export. Unavailable or unlicensed imagery must produce an explicit unavailable state, not a fabricated observation.

What the energy model compares

ARISOL_ROOF_MODEL_V0_1 compares baseline and proposed roof assumptions. It is an interpretable screening model, not a calibrated whole-building simulation. Color controls select reflectance assumptions; they do not measure a roof from RGB pixels.

The physical starting point is the change in absorbed solar energy: roof area × incident solar energy per area × change in solar reflectance. A shading adjustment and assumptions about heat transfer, cooling operation and HVAC efficiency are needed to translate a surface-energy change into cooling electricity. Reflectance and thermal emittance are different physical properties.[6]

Inputs: area (m²), annual solar energy (kWh/m²), peak solar irradiance (W/m²), baseline/proposed reflectance and emittance, insulation (RSI), HVAC COP, cooling and shade fractions, electricity rate, cost per ft², climate period and building use.
Outputs: conservative/base/optimistic annual cooling electricity (kWh/year), peak reduction (kW), electricity cost savings ($/year), and payback when meaningful cost and savings inputs exist.
Interpretation: absorbed solar reduction is not electricity saved. Peak reduction is not a guaranteed utility demand-charge reduction. Simple payback is not discounted cash flow or a contractor quote.

Emittance changes are not modeled in this version. Baseline and proposed emittance are recorded for transparency; the simplified heat-transfer approach assumes comparable high-emittance roofs. Low or differing emittance requires detailed simulation. The conservative/base/optimistic cases vary a fixed exterior heat-transfer coefficient rather than reporting measured uncertainty. No benefit is credited when reflectance does not improve; possible increased loads are not quantified.

Scenario ranges express sensitivity to assumptions; they are not statistical confidence intervals unless independently calibrated as such. Actual results depend on insulation, occupancy, HVAC operation, roof condition, weather, tariffs and installation. Heating penalties, moisture risk, maintenance, aging and replacement timing need a project-specific assessment. No Arisol material performance is validated by choosing a high-reflectance scenario.

Top Leads: coarse to fine, not every roof at once

Bound the territory by a center and radius; apply environmental screening where available.
Collect eligible candidate buildings from a permitted source.
Apply lower-cost filters and preliminary scoring.
Enrich only a bounded shortlist with building data and permitted imagery.
Rank qualifying results using transparent opportunity factors and confidence; return up to ten.

The 0–100 opportunity score is a prioritization heuristic, not validated machine-learning accuracy, an owner’s intent to buy or a probability of closing a sale. Inspect factor contributions and data gaps. Sparse coverage can produce fewer than ten results; that is preferable to inventing prospects. Radius limits and enrichment budgets protect latency and provider costs.

The validation we still need

Software tests can verify units, monotonic behavior, input validation, radius calculations and ranking. They cannot prove real energy savings. Field validation should pair roof measurements and installation records with weather-normalized pre/post energy data, record uncertainty, and use held-out buildings before publishing accuracy claims.

For materials, repeat batches and matched control samples should record substrate, thickness, curing, instrumentation, weather, solar reflectance, thermal emission and aging. A surface-temperature comparison alone does not establish whole-building savings, long-term durability or suitability for an occupied roof.

Read the science · Discuss a validation partnership

A screening estimate. Not a site inspection.