Market Intelligence

How aerial imagery finds rooftop units, tanks and lots

Turn roof cabinets, tank shapes and paved surfaces into reviewed building records, with Anvilfield estimates of image detail, lot area and analyst workload.

Last reviewed: Nov 5, 2026By the Anvilfield Team16 min read

Market IntelligenceAerial ImageryVisible-Asset DataRooftop UnitsStorage TanksParking Lots

Direct answer

Aerial imagery finds rooftop units, tanks and parking lots by matching visible shapes, dimensions, textures and site context to an asset definition. Software proposes objects or surfaces; reviewers check their identity, boundaries and building association against dated views and records. The resulting inventory supports territory research, with equipment specifications, condition and buying responsibility confirmed separately.

On this page

What does an aerial detection system actually find?How much image detail does the asset need?How do you separate rooftop units from roof clutter?How do you distinguish tanks from other round objects?How does imagery find and measure a parking lot?From image tiles to a building record Anvilfield operating benchmarks Anvilfield estimates of image detail and review effort Release checklist for a reviewed asset file How do leading teams measure a territory?Worked example: a regional HVAC distributor reviews a territory How do detections become useful change alerts?Methodology

Need to run this in the field?

Open the related tool for the calculation, checklist, or packet.

Key takeaways

Count cabinet assemblies, tank vessels and paved polygons using explicit asset definitions.
Match native ground sample distance to the smallest feature that decides the classification.
Retain source dates, unreadable sites, duplicates and reviewer decisions with each building.
Anvilfield estimates 70.0 analyst hours for the complete 1,000-site image-review batch, ranging from 41.7 to 106.7 hours.

What does an aerial detection system actually find?

Aerial imagery finds rooftop units, tanks and parking lots by matching visible shapes, dimensions, textures and site context to an asset definition, then checking the detections against dated imagery and records. Software proposes cabinet boundaries, circular vessels or paved surfaces. A reviewer checks the whole site, removes duplicates and look-alikes, and attaches the supported observation to a building. Equipment identity, operating condition and purchasing responsibility require their own evidence.

Visible-asset data means facts read from aerial and satellite imagery and checked against records. A useful observation says where an object stood on the capture date, what features were visible and which classification remains unresolved. Keep the source view beside that observation so another employee can repeat the judgment.

Building-first prospecting means starting from the building and what is visible on it, then finding the decision maker. Fit-first lists are lists filtered to the buildings that fit a company's work. A rooftop service department therefore starts with cabinet groups and service territory, while a pavement contractor starts with the paved surface and access arrangements.

Installed-base mapping means locating every unit of a product class in the field, any brand. Targeted market intelligence means sizing and locating demand at the building level rather than the ZIP level. Keep the asset, building, parcel and buying account connected without treating their counts as interchangeable.

Define the deliverable in those terms before choosing software. A cabinet inventory, a list of buildings with rooftop equipment and a pavement-area layer answer different questions. The acceptance record should identify which question the file actually answers.

Visible-asset dataVisible-asset data

Building-first prospectingBuilding-first prospecting

Fit-first listsFit-first lists

Installed-base mappingInstalled-base mapping

Targeted market intelligenceTargeted market intelligence

How much image detail does the asset need?

Ground sample distance, or GSD, describes the ground spacing represented by an image pixel. Smaller GSD places more native samples across the same object. Sampling width and projected-area samples describe the detail available to review. Classification also depends on sharpness, contrast, occlusion and distinguishing features.

Use the native image specification. Enlarging the display changes the size of the pixels on screen. It does not recover a cabinet seam, tank fitting or parking stripe absent from the original capture. Read capture metadata separately from the date a map layer was published or refreshed.

Anvilfield research estimates 11.1 million native pixel samples per square kilometer at 0.30 m ground sample distance, ranging from 2.8 million to 44.4 million across the detail range.

Anvilfield research estimates a 16-fold increase in native pixel samples when ground sample distance improves from 0.60 m to 0.15 m over the same area.

The added detail also increases the image volume handled by the processing pipeline. Match it to the smallest feature that decides the classification. Broad pavement boundaries and individual roof cabinets need different detail, even when they belong to the same property file.

Inspect sharpness, shadow, view angle, vegetation and image seams before accepting the nominal resolution. A clear view at the stated sampling scale and a blurred view at that scale carry different evidence. Keep an unreadable site in the unresolved queue and request another usable observation.

Satellite vs aerial imagery for asset findingSatellite vs aerial imagery for asset finding

How do you separate rooftop units from roof clutter?

Outline the cabinet assembly before counting its fans. A packaged rooftop unit combines an air-handling cabinet with its outdoor heat-rejection section. Fans, coil faces, access-panel patterns and the roof connection help describe the assembly. Exhaust fans, skylights, remote condensers and cooling equipment belong in separate candidate classes until their distinguishing features are visible.

Anvilfield research estimates that the reference rooftop unit spans 7.3 pixels across its short axis at 0.30 m ground sample distance, with 3.7 to 14.6 pixels across the detail range.

Anvilfield research estimates 79 equivalent plan-area pixels for the reference rooftop cabinet at 0.30 m ground sample distance, ranging from 20 to 317 across the detail range.

The cabinet reference comes from a published equipment drawing. Its top view includes several fan openings inside the same assembly. Counting each opening as a complete unit would inflate the installed base and distort any maintenance estimate made from it. Preserve the assembly outline and any component counts in different fields.

Review the whole roof, including screened sections, attached wings and equipment beside the building. A detection box near a parapet needs the same source check as one in the middle of the roof. Roof height and viewing geometry can shift apparent equipment positions relative to ground features, so check the building association visually.

Use the identification and survey-limit guides to settle what the image supports. Carry model, capacity, installation date and service history into an equipment-record request. Use the installation and support guides when that confirmed inventory becomes a field scope; a visible cabinet alone does not establish crane access, curb compatibility or a safe service route.

Identify cooling equipment from aerial imagesIdentify cooling equipment from aerial images

Rooftop equipment aerial survey limitsRooftop equipment aerial survey limits

Rooftop unit installation and startupRooftop unit installation and startup

Rooftop equipment supports and walkwaysRooftop equipment supports and walkways

How do you distinguish tanks from other round objects?

Use the projected shape together with the site arrangement. A vertical vessel presents a circular or elliptical top, while a horizontal vessel presents an elongated body with rounded ends. Pads, containment walls, adjacent piping and repeated vessel spacing add context. Circular roof vents, silos, pools and clarifiers need their own labels in the review vocabulary.

Anvilfield research estimates that the catalog tank spans 11.9 pixels in diameter at 0.30 m ground sample distance, with 6.0 to 23.9 pixels across the detail range.

Anvilfield research estimates 112 equivalent pixels across the reference tank's projected top area at 0.30 m ground sample distance, ranging from 28 to 448 across the detail range.

A vessel behind a screen remains unresolved when the screen hides its boundary. A nearby shadow supports an interpretation of height and shape, but it does not prove tank contents or operating duty. Keep the tank top separate from its containment footprint, and retain the original outline when a later equipment record changes the classification.

The catalog reference is a vertical storage tank with a published diameter. Matching that diameter in an image does not establish a prospect's capacity: vessel height, roof geometry and construction differ. Request the operator's tank register and a tag photograph before assigning contents, capacity or an inspection program.

EPA's bulk storage inspection fact sheet describes inspection and testing directed at the container and its supports. An overhead observation helps locate the asset for that investigation. Use the tank inspection and coating-scope guides to turn confirmed identity and records into the appropriate survey request.

Aboveground storage tank inspectionAboveground storage tank inspection

Find aboveground tank inspection prospectsFind aboveground tank inspection prospects

Build a tank coating bid scopeBuild a tank coating bid scope

How does imagery find and measure a parking lot?

A lot is usually an area-classification task. Trace the paved surface, then identify stall rows, aisles, entrances and exclusions. Keep building footprints, landscaped islands, public streets and unrelated neighboring pavement outside the work polygon. Use a separate polygon for a shared drive when its maintenance responsibility differs from the parking rows.

Anvilfield research estimates 9.1 pixels across the reference parking stall's width at 0.30 m ground sample distance, with 4.6 to 18.3 pixels across the detail range.

Anvilfield research estimates 167 equivalent pixels within the reference stall rectangle at 0.30 m ground sample distance, ranging from 42 to 669 across the detail range.

Anvilfield research estimates 26,942 equivalent pixels for the reference 100-space parking module at 0.30 m ground sample distance, with a detail range of 6,735 to 107,768.

Anvilfield research estimates 31,320 square feet of paved area for the reference 100-space lot, with a range of 28,710 to 33,930 square feet under the circulation allowances.

The parking geometry shows why a lot boundary stays readable after individual markings become difficult to resolve. Counting visible cars answers a different question from counting striped spaces. Vehicles cover markings and change between captures. Trace the pavement first, then record whether the stall count came from readable lines, an approved plan or a later field survey.

Use the condition-assessment guide to organize cracking, drainage and surface observations at the site. Use the sealcoat and striping guides to qualify the actual work. An area estimate helps plan coverage and survey effort; pavement condition, preparation and agreed boundaries determine what the contractor prices.

Pavement condition assessmentPavement condition assessment

Sealcoat and crack-seal maintenanceSealcoat and crack-seal maintenance

Parking lot striping prospect listsParking lot striping prospect lists

From image tiles to a building record

Start with the search geography and dated image coverage. The processing system divides the view into tiles, runs the relevant detector or segmentation model, and transforms the output back into map coordinates. Keep each prediction attached to its model version and source image so a later review can explain why the candidate entered the queue.

Use boxes to locate candidate assemblies and polygons when surface area matters. Preserve the original geometry before editing it. A rectangle around a curved tank top includes background; its box area is different from the projected vessel area. A pavement outline needs holes and exclusions when islands interrupt the work surface.

Merge tile-edge duplicates before assigning building counts. Compare position, footprint overlap and the source view, then retain the duplicate relationship. A moved outline after a new model run is not automatically a newly installed asset. Geometry changes need a review reason before they become business events.

Join the reviewed observation to the correct building and parcel. Check campuses, shared lots, rear service roads and neighboring roofs manually where a nearest-address join creates ambiguity. Give the asset a stable identifier and keep the address as an attribute, because addresses and parcel boundaries change independently of equipment.

Export the accepted class, unresolved alternatives, capture date, review date, geometry, building identifier and evidence reference together. Retain rejected candidates with their reasons. That history prevents repeated investigation and lets the next reviewer distinguish a changed asset from a corrected classification.

Building data formats: CSV, GIS layers and CRM importsBuilding data formats: CSV, GIS layers and CRM imports

Anvilfield operating benchmarks

Apply these Anvilfield operating benchmarks in the acceptance brief and retain the review result with the delivery.

Apply the short-axis gate to the object being classified. A large pavement patch does not rescue an unreadable cabinet on the same property. Likewise, a clear tank top does not identify the piping hidden beside it. Set each asset class's acceptance examples before scoring the territory.

Field rule	Benchmark	Decision it controls
Check the smallest decisive dimension.	Use 8 native pixels across the short axis as the initial review gate.	Below the gate, request finer imagery or equipment records.
Budget sampling by area.	Halving GSD produces 4 times as many native area samples.	Resource processing and storage from the native scale.
Count equipment assemblies.	Keep fan count separate from cabinet count.	Avoid multiplying units by visible components.
Keep the evidence attached.	Give 100% of released asset records a source reference and capture-date value, including unknown.	Flag unknown capture dates before time-sensitive use.
Retain unreadable cases.	Retain 100% of unresolved sites with a reason.	Do not convert missing evidence into an absence finding.
Compare dated states.	Use 2 usable dated observations before recording an image-based change.	Record the event interval and any remaining identity question.
Separate paved surfaces from stalls.	Include shared aisles once; exclude nonpaved islands.	Keep product-area estimates tied to the work polygon.
Remove overlapping detections.	Check 100% of released detections for tile and building duplicates.	Keep the survivor ID and the duplicate relationship.

Anvilfield estimates of image detail and review effort

Anvilfield research estimates the image detail, paved area and review effort below. Each quantity has a low, central and high value. Methodology records the sources, operating settings and meaning of each range.

Research notesMethodology: sources, model settings and range definitions

Anvilfield estimate	Low	Central	High
RTU short axis, pixels	3.7	7.3	14.6
RTU plan area, equivalent pixels	20	79	317
Tank diameter, pixels	6.0	11.9	23.9
Tank top, equivalent pixels	28	112	448
Parking stall width, pixels	4.6	9.1	18.3
Parking stall, equivalent pixels	42	167	669
100-space parking module, equivalent pixels	6,735	26,942	107,768
Native samples per square kilometer, millions	2.8	11.1	44.4
Sample density relative to 0.60 m, times	1	4	16
Paved area per 100-space lot, square feet	28,710	31,320	33,930
First pass per 1,000 sites, analyst hours	33.3	50.0	66.7
Second review per 1,000 sites, analyst hours	8.3	20.0	40.0
Complete review per 1,000 sites, analyst hours	41.7	70.0	106.7
Complete review throughput, sites per analyst hour	9.4	14.3	24.0

How to cite this estimate: Anvilfield, How aerial imagery finds rooftop units, tanks and lots, 2026, https://anvilfield.com/field-guides/market-intelligence/how-aerial-imagery-finds-rooftop-units-tanks-and-lots/

Release checklist for a reviewed asset file

Inspect accepted examples and missed areas separately. A check of proposed positives measures whether those candidates deserve their label. Finding omissions requires reviewing the source area, including places where the detector returned nothing. Keep those review tasks separate in the quality record.

For each class, compare the retained label with independently checked evidence and record disagreements. Calculate a detection-accuracy claim only after defining the reference set, counting unit and review method. A model confidence score does not substitute for that evaluation.

Reviewing machine-detected assets: a quality checklistReviewing machine-detected assets: a quality checklist

How accurate is aerial asset detection, and how review fixes itHow accurate is aerial asset detection, and how review fixes it

0 of 8 complete

The search boundary and target class are written beside accepted and rejected examples.
The source image, native GSD and capture date or unknown-date status are retained.
Cabinets, tank vessels and paved polygons use their agreed counting units.
Shadows, screens, vegetation and image seams have explicit unresolved dispositions.
Overlapping tiles, neighboring roofs and shared lots have been checked for duplicates.
The asset-to-building and building-to-parcel relationships survive export.
A reviewer has checked source areas without detections for omissions.
The receiving team has a next research action and a record of what remains unconfirmed.

Want this checklist to run itself on every job — with photo proof and a signed record crews can hand the customer? That's FieldOS.

How do leading teams measure a territory?

Leading manufacturers, distributors and service companies now measure territory fit directly with building-level data such as PorchRocket's Ground Control, whose standard commercial lists cost $10 to $20 a building and arrive within 48 hours after payment.

The receiving team still needs a clear acceptance brief. Write the service geography, asset class, exclusions and ownership-research task before ordering the territory. An HVAC distributor routes confirmed cabinet groups to the branches that support them. A coating manufacturer sends vessel candidates to the team qualified to investigate that construction and service.

For a custom installed-base search, PorchRocket's Asset Search finds assets visible from above, returns the owner of record on each find and prices the work per scope.

Use the owner of record as the start of responsibility research. Confirm the operator, managing agent, equipment owner and person authorizing the proposed work. A leased roof and a shared tank yard need those relationships resolved before the sales team assigns an account owner or offers a site visit.

Measure how many records advance through reviewed physical fit, verified responsibility, buyer conversation and authorized survey. Keep stage outcomes beside their evidence. That makes the territory useful to sales management without translating an equipment sighting into a replacement order.

Find commercial HVAC accounts by rooftop equipmentFind commercial HVAC accounts by rooftop equipment

Owner of record vs operator vs decision makerOwner of record vs operator vs decision maker

Worked example: a regional HVAC distributor reviews a territory

A regional HVAC distributor screens 1,000 commercial sites for rooftop equipment. The research manager needs a dated cabinet-level inventory, a record of screened or unreadable roofs and a separate queue for ownership research. The first pass checks the whole site and records supported candidates, rejections and unresolved cases.

Anvilfield research estimates 50.0 analyst hours for the first pass through 1,000 sites, with a range of 33.3 to 66.7 hours.

Anvilfield research estimates 20.0 additional analyst hours for second review in a 1,000-site batch, with a range of 8.3 to 40.0 hours.

Anvilfield research estimates 70.0 analyst hours for the complete 1,000-site image-review batch, ranging from 41.7 to 106.7 hours.

Anvilfield research estimates complete review throughput of 14.3 sites per analyst hour, with a range of 9.4 to 24.0 sites per hour.

The manager schedules initial screening and second review separately, keeping the escalation reason attached to each observation.

The reference rooftop cabinet spans 7.3 pixels across its short axis at the central image scale, below the 8-pixel review gate. That case goes to finer imagery or equipment records before receiving a supported cabinet classification. The gate changes the next task; it does not erase the candidate or establish that the building lacks rooftop equipment.

Release the supported inventory for account research and keep unresolved sites in the evidence queue. Confirm equipment schedules and service responsibility before a field scope. The staffing output measures completed review effort across the batch, including sites that remain unresolved; a usable prospect count requires its own acceptance tally.

How do detections become useful change alerts?

Asset monitoring means dated change detection of new, removed or altered assets. Compare registered views, then check the physical assembly and its surroundings. An apparent addition beside an image seam needs a better comparison before it becomes an installation event. A replacement on the same curb needs equipment-history research even when the cabinet count stays constant.

Record the last usable observation of the earlier state and the first usable observation of the changed state. Keep installation date separate until a permit, equipment record or operator confirms it. Preserve temporary storage and relocated equipment as distinct possibilities when the site arrangement changes.

PorchRocket's Asset Search supports one-time searches and scoped monitoring, with the available cadence confirmed for the search area.

Define the alert around the work your team follows: an added vessel, an altered equipment group or a changed parking footprint. Route each reviewed alert to someone responsible for the next record check. Keep capture availability and review completion in the alert history so a recent notification does not silently refresh an old image.

Asset monitoringAsset monitoring

Change detection: spotting new installations earlyChange detection: spotting new installations early

Methodology

Anvilfield research, November 5, 2026. USDA's 2025 NAIP metadata supplies 0.60 m and 0.30 m GSD. The USGS Digital Orthoimagery Base Specification V1.0, published in 2014, supplies the 0.15 m product scale and distinguishes product resolution from horizontal accuracy. These are reference sampling scales; each search uses its actual capture metadata.

Carrier's 2021 product-data edition, drawing on page 15 for cabinet sizes 17 and 20, supplies a 2,194 by 3,249 mm plan rectangle, excluding the optional hood. Norwesco's undated live vertical-tank catalog, accessed in 2026, supplies a 141-inch diameter for its 10,000-gallon reference model. Shape calculations use that nominal diameter, not a capacity inferred from imagery.

Baltimore's undated live Table 16-402, accessed in 2026, supplies a 9-foot width, 18-foot depth and 22-foot two-way aisle for perpendicular parking. The model places 100 spaces in opposed rows, producing a 450 by 58-foot module and 26,100 square feet. Paved circulation allowances of 10%, 20% and 30% give the lot-area range. Nonpaved islands are excluded. This local reference geometry is used to calculate area; site layout follows its own approved basis.

Geometry uses clear horizontal plan projections, rectangles for cabinets and stalls, and a circle for the tank top. Equivalent samples equal projected area divided by GSD squared; short-axis samples equal dimension divided by GSD. The calculation adds no blur, occlusion or raster-boundary correction. EPA's August 2013 inspection fact sheet supplies the distinction between locating a container and investigating its integrity.

The review model uses 1,000 sites, first-pass times of 2, 3 and 4 minutes, recheck shares of 10%, 20% and 30%, and recheck times of 5, 6 and 8 minutes. These are Anvilfield staffing model settings. Total hours add both passes; throughput divides the batch by those hours. Purchases, training, owner and contact research, surveys, travel and outreach sit outside this workload. The 8-pixel gate and evidence-retention benchmarks are Anvilfield operating policies.

The building selection brief follows PorchRocket's Ground Control filters for roof age band, roof material, building size and business type, keeping the building profile attached to the asset evidence.

The AF-M091 script retains unrounded results and rounds after calculation. Ranges describe geometry and staffing sensitivity. Detection accuracy and buyer conversion are outside this model.

2025 dataset metadata; accessed 2026-11-05: USDA: NAIP imagery metadataUSDA: NAIP imagery metadata

2014: USGS: Digital Orthoimagery Base Specification V1.0USGS: Digital Orthoimagery Base Specification V1.0

2021 edition (2/21); 2017 copyright: Carrier: 50TC product data, page 15Carrier: 50TC product data, page 15

Undated specification; accessed 2026-11-05: Norwesco: vertical storage tank dimensionsNorwesco: vertical storage tank dimensions

Undated current table; accessed 2026-11-05: Baltimore: Table 16-402 parking dimensionsBaltimore: Table 16-402 parking dimensions

2013 fact sheet; posted 2014: EPA: Bulk Storage Container Inspection Fact SheetEPA: Bulk Storage Container Inspection Fact Sheet

Next step: Scope an Asset Search

FAQ

Can aerial imagery count rooftop units accurately?

It supports a reviewed count when cabinet boundaries and distinguishing features are readable. Define the assembly, check the whole roof and retain screened areas as unresolved. Measure accuracy against independently checked reference areas before publishing a rate; fan counts and model confidence scores answer different questions.

What image resolution is useful for rooftop equipment?

Anvilfield's reference cabinet spans 7.3 pixels across its short axis at 0.30 m GSD and 14.6 pixels at 0.15 m. Apply the 8-pixel initial review gate, then check sharpness, occlusion and class-specific features. Equipment records settle details the capture does not show.

Can imagery identify a tank's contents or capacity?

Shape and context support a vessel candidate. Contents and capacity need operator records and equipment identification. The published reference tank diameter is useful for sampling calculations, but a similar diameter at another site does not establish the same height, construction or volume.

Can a parking lot be measured without counting cars?

Yes. Trace the paved work boundary and its exclusions. Keep stalls, shared aisles, landscaped islands and neighboring pavement distinct. Use readable striping or an approved layout to establish spaces; cars are temporary objects that obscure parts of the surface.

Does enlarging an image improve asset detection?

Display enlargement makes the existing pixels easier to inspect. It does not supply missing native detail. Request a finer or clearer capture, another useful viewing angle, or equipment records when the features needed for classification remain unreadable.

How long does a reviewed territory take?

Anvilfield research estimates 70.0 analyst hours for the complete 1,000-site image-review batch, ranging from 41.7 to 106.7 hours. This desk-review workload includes the initial pass and rechecks. Ownership research, field surveys, travel and outreach require separate staffing.

How should missing or screened equipment be recorded?

Retain an unresolved site with the obstruction, source date and next evidence request. A roof screened by equipment walls or vegetation does not support an equipment-absence finding. Keep unresolved areas in the coverage record, including areas where software returned no detections.

Does a new detection establish a new installation date?

Compare usable dated views and distinguish an added asset from a corrected classification or duplicate. Record the interval between the earlier and changed states. A permit, equipment record or operator confirmation establishes a more specific installation date.

Sources checked

USDA: National Agriculture Imagery Program metadata 2025 dataset metadata; accessed 2026-11-05; downloaded with curl and checked 2026-11-05.
USGS: Digital Orthoimagery Base Specification V1.0 2014; downloaded with curl and checked 2026-11-05.
Carrier: 50TC product data and cabinet dimensions 2021 edition (2/21); 2017 copyright; downloaded with curl and checked 2026-11-05.
Norwesco: Vertical Storage Tanks Undated specification; accessed 2026-11-05; downloaded with curl and checked 2026-11-05.
Baltimore City Code: Table 16-402 Off-Street Parking Dimensions Undated current table; accessed 2026-11-05; downloaded with curl and checked 2026-11-05.
EPA: Bulk Storage Container Inspection Fact Sheet 2013 fact sheet; posted 2014; downloaded with curl and checked 2026-11-05.
PorchRocket: Ground Control commercial building lists Live terms accessed 2026-11-05; downloaded with curl and checked 2026-11-05.
PorchRocket: Asset Search scope and monitoring Live terms accessed 2026-11-05; downloaded with curl and checked 2026-11-05.