Market Intelligence
How accurate is aerial asset detection, and how review fixes it
Measure precision and recall separately, audit missed assets, and use a sourced review model to plan a reliable building-level inventory.
Direct answer
Aerial asset detection is useful when precision and recall are measured separately and reviewers check both flagged objects and unflagged coverage. Anvilfield estimates reviewed cooling-tower list precision at 97.6%, ranging from 93.2% to 99.5%, and visible-inventory recall at 94.9%, ranging from 87.6% to 98.3%. Candidate cleanup removes wrong records; a separate coverage search recovers missed assets. Confirm equipment identity and service scope before quoting work.
On this page
What does accurate mean for an asset list?Why do aerial detections get the wrong answer?How much does review improve detection accuracy?How do you test a detector without hiding missed assets?Which review rules belong in the benchmark?How does review fix the list and the inventory?Worked example: a regional cooling-tower service companyWhat should leaders put in an asset-data scope?How do you keep the review useful for sales?Release checklistHow do you keep accuracy from drifting?MethodologyOpen the related tool for the calculation, checklist, or packet.
Key takeaways
- Precision measures the correctness of returned records; recall measures how much of the real target inventory was found.
- Anvilfield estimates reviewed cooling-tower precision of 97.6% and visible-inventory recall of 94.9%.
- An independent audit must include unflagged coverage to measure missed assets.
- Keep capture dates, review dates, asset identities and corrections with the released file.
What does accurate mean for an asset list?
Ask which claim the accuracy score describes. Finding equipment, assigning its class, counting separate assemblies, placing it on the right parcel and confirming the responsible buyer are different tasks. A useful acceptance report gives each task a denominator. A polished map with correct-looking pins leaves those questions open until someone checks the underlying records.
Precision measures how many returned candidates are correct. Recall measures how many real target assets the search found. A true positive is a correctly matched asset; a false positive is an incorrect candidate; a false negative is a real asset the search missed. Score individual assemblies against an independently established reference inventory. Count duplicate detections as extra false records after the valid match has been assigned.
Visible-asset data means facts read from aerial or satellite imagery and checked against records. Its accuracy depends on the particular observation being made. A clearly visible fan supports a fan observation, while a complete cooling-tower inventory requires assembly boundaries, equipment classification and coverage of the entire agreed search area.
Keep a model confidence score in its own column. A high score orders the review queue; an independently checked sample measures precision. Also separate asset-level recall from building-level recall. Finding equipment somewhere on a campus does not establish that every serviceable assembly on that campus has been counted.
Visible-asset dataRooftop equipment aerial survey limitsWhy do aerial detections get the wrong answer?
Most review problems start with a plausible resemblance. Fan banks, cooling equipment, skylights and roof furniture share outlines. A reviewer needs the surrounding connections and the whole assembly to decide what the shape represents. The CDC identification guidance and labeled photographs support comparison of tower configurations, including equipment whose fans sit on the side. [S2, S3]
Separate recognition errors from visibility failures. A visible dry cooler labeled as a tower is a classification error. A tower hidden behind screening is a coverage problem. Removing the first error improves list precision; resolving the second requires additional evidence. Enlarging the same image does not reveal an obscured connection or a hidden enclosure.
Boundary errors deserve their own rejection reasons. Adjacent fans inside one assembly should remain associated with that assembly. Overlapping image tiles should resolve to a stable asset identifier. A roof edge beside another parcel needs a location check before ownership is attached. A correct equipment class on the wrong property still sends the sales team to the wrong account.
Time introduces another failure mode. An image depicts its capture date. A recent review date tells you when a person interpreted it. Replacement records, current site photographs and authorized confirmation establish later facts. Keep these dates separate so a removed asset does not survive indefinitely in a supposedly current inventory.
Use the technical guides below to settle the classification and scope questions that matter to your trade. Cooling-equipment recognition establishes the likely family; commissioning and water-treatment records establish the installed system and its operating responsibilities. An image label is the start of that investigation.
S2: CDC, Strategies for Identifying Cooling Towers (2026)S3: CDC, Pictures of Cooling Towers (2024)Identify cooling equipment from aerial imagesCooling tower types and operationCooling tower commissioningHow much does review improve detection accuracy?
Anvilfield estimates the following review outcomes per 1,000 real, aerially visible cooling towers. List precision measures returned-record correctness; inventory recall measures the visible towers found across the search area.
Anvilfield estimates a central machine queue of 1,071 candidate records, ranging from 964 to 1,177, with 916 correct detections, ranging from 869 to 951. Anvilfield estimates 155 false positives, ranging from 95 to 226, and 84 missed towers, ranging from 49 to 131.
Anvilfield estimates review removes 132 false positives, ranging from 67 to 215, and loses 9 correct detections, ranging from 4 to 29. Anvilfield estimates the coverage search recovers 42 missed towers, ranging from 12 to 98.
Anvilfield estimates the reviewed list contains 949 correct tower records, ranging from 876 to 983, and 23 false positives, ranging from 5 to 68. Anvilfield estimates 51 real visible towers remain absent, ranging from 17 to 124.
Anvilfield estimates reviewed-list precision of 97.6%, with a range of 93.2% to 99.5%, and visible-inventory recall of 94.9%, with a range of 87.6% to 98.3%.
| Anvilfield estimate | Low | Central | High |
|---|---|---|---|
| Machine candidate records | 964 | 1,071 | 1,177 |
| Correct machine detections | 869 | 916 | 951 |
| False positives before review | 95 | 155 | 226 |
| Real towers missed before review | 49 | 84 | 131 |
| False positives removed | 67 | 132 | 215 |
| Correct detections lost in review | 4 | 9 | 29 |
| Missed towers recovered | 12 | 42 | 98 |
| Correct towers in reviewed list | 876 | 949 | 983 |
| False positives remaining | 5 | 23 | 68 |
| Real towers still absent | 17 | 51 | 124 |
| Reviewed-list precision | 93.2% | 97.6% | 99.5% |
| Reviewed inventory recall | 87.6% | 94.9% | 98.3% |
How to cite this estimate: Anvilfield, How accurate is aerial asset detection, and how review fixes it, 2026, https://anvilfield.com/field-guides/market-intelligence/how-accurate-is-aerial-asset-detection-and-how-review-fixes-it/
How do you test a detector without hiding missed assets?
Freeze the search boundary, eligible asset definition, evidence dates and matching rule before scoring. Build the reference inventory independently of the candidate list. A reviewer who sees only machine flags can discover incorrect flags but cannot establish what the detector omitted. Include the agreed roofs, yards and other eligible surfaces in the reference search.
Randomly sample accepted candidates to estimate precision. Separately select geographic areas for exhaustive reference review to estimate recall. Preserve the sample-selection method and inclusion probabilities. When you deliberately oversample difficult roofs or rare asset classes, report those strata separately or weight them back to the target population. A pooled score from a convenience sample hides where the work is weakest.
Keep an untouched validation batch. Resolve reviewer disagreements against stronger evidence, retaining the original decisions and the adjudication. Use field records or authorized site confirmation when imagery cannot settle the identity. Mark obscured areas as unavailable coverage, and report that coverage alongside the visible-asset score. This keeps the denominator tied to what was actually evaluated.
Anvilfield calculates that an audit with 95 correct records in 100 random checks has a 95% Wilson interval of 88.8% to 97.8%. At 190 correct records in 200 checks, Anvilfield calculates 91.0% to 97.3%; at 380 in 400, it calculates 92.4% to 96.7%. The point estimate stays 95%, while the uncertainty narrows. [S4]
Those intervals use independent random binary checks. Roofs from the same campus and repeated views of the same assembly are related observations. Design the sample around those clusters instead of counting every crop as independent evidence. Freeze the audit before reviewing it; tuning the model or the acceptance rule against the audit turns that batch into development data.
S4: NIST/SEMATECH, Confidence intervals (undated; accessed 2026)S6: Leber, Pibida and Enders, NIST Technical Note 2045 (2019)Reviewing machine-detected assets: a quality checklistWhich review rules belong in the benchmark?
Use these Anvilfield benchmarks when buying or building a reviewed inventory. Set the acceptance rule against the decision: a territory-screening file and a complete equipment-service schedule need different proof.
A clean audit is useful only for the claim sampled. Random accepted-record checks test the correctness of the delivered rows. They leave the unsearched territory and hidden equipment outside their conclusion. The exact binomial calculation below handles a zero-error sample without declaring the population error-free. [S5, S6]
S5: NIST Dataplot, Exact Binomial Confidence Limits (2010)| Field rule | Benchmark | Why it changes the decision |
|---|---|---|
| Count the asset once | Use one matched record per assembly; retain components underneath it. | Unit counts, parcel joins and duplicate handling use the same identity. |
| Publish both metrics | Report precision and recall separately for each asset class. | A cleaner candidate list does not establish full coverage. |
| Keep the evidence clocks | Record capture date, review date and confirmation date separately. | An old view does not become current when somebody opens it. |
| Audit independently | Use random candidate checks for precision and independent area searches for recall. | The machine output cannot serve as its own complete reference. |
| Show uncertainty | At 190 correct results in 200 checks, report 95% with a 91.0% to 97.3% Wilson interval. | A point estimate alone overstates what the sample establishes. |
| Predeclare the acceptance rule | 59 independent checks with zero errors support an error rate below 5% at 95% one-sided confidence. | Select the fixed sample in advance; do not keep sampling until the record looks clean. |
| Keep unresolved work visible | Give each unresolved record a reason, an evidence owner and a next check. | Removing difficult cases from the report conceals the remaining research burden. |
| Measure reviewer losses | Log retained true detections, rejected true detections and recovered omissions separately. | Candidate cleanup and missed-asset recovery require separate checks. |
How does review fix the list and the inventory?
Run a candidate-review pass and a coverage-review pass, with separate results. Candidate reviewers examine the original view and surrounding context, merge duplicates, correct locations and reject look-alikes. Coverage reviewers search the eligible areas that received no candidate, including image boundaries and confusing installations. Give them access to the same asset definition but preserve an independent reference judgment.
Record what each correction changed. A class correction changes the product family. A parcel correction changes the recorded property. A duplicate removal changes the count. An evidence-date correction changes whether the record is ready for outreach. A reason such as bad data combines those distinct problems and prevents the next delivery from improving.
Reviewers also make mistakes. Keep disputed genuine assets in a reconciliation queue and measure how often acceptance rules discard them. A strict filter improves the appearance of the list while reducing inventory recall if it removes difficult true detections. The central calculation explicitly includes that loss before adding independently corroborated omissions.
For custom territory research, PorchRocket's Asset Search provides reviewed results with the owner of record for assets visible from above. Put the asset definition, unresolved-case treatment and missing-coverage review in the search scope so the delivery can be checked against the same claims your team intends to use.
Finish with a versioned correction ledger. Preserve the original asset ID, old value, new value, reviewer, evidence and disposition. Carry accepted corrections into the next release, and keep rejected records searchable so they are not repeatedly rediscovered. Review pays back when its findings change the file and the next search, rather than ending in an isolated spreadsheet.
RTU install and startup field guideWorked example: a regional cooling-tower service company
A regional cooling-tower service company prepares a territory inventory for its maintenance sales team. Anvilfield's central case starts with 1,000 real, visible towers. The operations manager separates asset identification from service qualification, assigns candidate and coverage queues, then commissions a local audit before releasing customer records.
The machine queue contains about 1,071 candidate records: 916 correct detections and 155 false positives. The reference population contains another 84 towers absent from that queue. If the company reviews only the flagged records, it can clean the contact list while leaving those omissions untouched. That distinction determines how the research work is assigned.
Candidate review removes 132 false positives and accidentally excludes 9 genuine detections. A separate coverage search recovers 42 initially missed towers. Anvilfield puts the completed list at 949 correct tower records and 23 false positives, with 51 real visible towers still absent. Precision reaches 97.6%; recall reaches 94.9%.
Without the coverage search, central recall is 90.7%. This gives the service manager a direct operating choice: fund only the cleanup when the task is reducing wrong calls, or also fund the missing-asset search when the task is understanding territory coverage. The same result count cannot answer both questions.
The sales team then groups accepted assets by building and service area. It confirms the operator, maintenance responsibility, access and equipment identity before proposing work. Several tower records at the same property belong to one coordinated account conversation. A reviewed sighting supports that conversation; equipment records and an authorized survey establish the service scope.
Service territory planningWhat should leaders put in an asset-data scope?
Building-first prospecting means starting with the building and its visible features, then finding the decision maker. Leading a manufacturer, distributor or service team through that process starts with a clear acceptance sheet: which asset classes qualify, how assemblies are counted, which locations are covered and what evidence each delivered row must preserve.
Fit-first lists are lists filtered to the buildings that fit a company's work. Write those filters before collecting contacts. Building use, equipment family, service geography, roof access and purchasing responsibility answer different questions. A matching building enters the research queue; the operator or authorized representative supplies the commercial and access facts needed to move it forward.
For a standard rooftop-equipment prospect file, PorchRocket's Ground Control filters buildings by visible rooftop HVAC, building size and business type, with CSV and Excel delivery within 48 hours after payment. Manufacturers and distributors can use the building file to define the territory, then audit the fields that control product fit and handoff to their channel partners.
Ground Control by PorchRocket lists Subscribe & save tiers of $10 to $20 per building on its rooftop HVAC page. Keep that procurement unit separate from equipment count: the budget buys building records, while the service opportunity depends on the assemblies and work confirmed at those buildings.
Specify the evidence package as carefully as the columns. Require the search boundary, observation dates, class definitions, review disposition and a way to reproduce the location. Agree who adjudicates disputes and how corrected records are returned. Request a sample that includes difficult cases and unflagged areas, because a hand-picked collection of easy roofs does not test the workflow you are buying.
Building-first prospecting; Fit-first listsHow do you keep the review useful for sales?
Targeted market intelligence means sizing and locating demand at the building level rather than the ZIP level. Accuracy matters because each wrong classification consumes research and outreach effort, while missed buildings leave the territory estimate incomplete. Keep the operational question beside the metric: which buildings should receive a qualification call, and which equipment families remain undercounted?
Installed-base mapping means locating units of a product class in the field, across brands. Maintain building and asset identifiers separately so an equipment replacement, a new operator or a corrected parcel does not create a duplicate account. Keep detected equipment, confirmed equipment and quoted work in separate stages of the commercial file.
Choose the technical follow-up from the asset. A roof record leads to an inspection and maintenance review. A tank candidate leads to identification, access and coating-scope questions. A parking-area sighting leads to pavement condition and maintenance-priority checks. Use the linked guides to turn the research record into a task an estimator or service manager can actually perform.
Keep error feedback close to the sales team. A caller should be able to mark wrong building, wrong equipment class, duplicate account or wrong responsible party without deleting the source observation. Route the correction to the person who owns that field. Track detection precision separately from contact reachability, appointment acceptance and completed surveys.
Targeted market intelligence; Installed-base mappingRoof maintenance programRoof asset managementTank coating bid scopeParking lot survey priorityRelease checklist
Release the file when the research owner can answer the following questions from the saved evidence. Keep unresolved cases in the delivery with an assigned follow-up. A reviewer should be able to reproduce the release decision without asking the original analyst what a confidence label meant.
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 you keep accuracy from drifting?
Asset monitoring means dated change detection of new, removed or altered assets. It starts with a baseline whose observation date, location and identity can be reproduced. Compare corresponding areas and preserve the interval between the views. Different visibility or equipment screening needs review before a changed outline becomes an installation or removal claim.
Recheck performance when the territory, asset mix, imagery characteristics or acceptance rules change. Keep recent audit results by asset class and evidence cohort. If a new delivery contains more unresolved roofs, examine the coverage before adjusting the confidence cutoff. A threshold change alone does not repair the missing evidence.
Make confirmed changes useful across the account. Preserve the prior asset record, attach the replacement or removal evidence and notify the internal research owner through the existing workflow. Feed the confirmed event into territory planning and service qualification. Keep the observed change date interval distinct from any installation date supplied later by the operator.
Asset monitoringHow fresh does building data need to be for salesMethodology
Research notes, November 5, 2026. S1 is the 2024 Wong and colleagues study of automated cooling-tower detection, with initial imagery collection in 2021. Its New York City and Philadelphia results report 90.1% precision and 95.1% recall; Boston reports 80.8% and 91.6%; Athens reports 85.5% and 86.9%. The saved script extracts those values from the downloaded author manuscript. The archive release in 2025 does not change the study year.
The sensitivity model uses the componentwise medians as its central machine inputs. Review outcomes are Anvilfield estimates computed from those inputs and the assumptions below. The study supplies measured machine baselines; rejection, retention and recovery rates are Anvilfield modeling choices. Applying the workflow to another equipment class starts with that class's own reference audit.
For reference population N, recall r and precision p, correct machine detections equal N x r; false positives equal correct detections x (1 - p) / p; omissions equal N minus correct detections. Reviewed true detections equal retained detections plus corroborated recoveries. Reviewed false positives equal original false positives x (1 - rejection fraction). Divide by the accepted candidate total for precision and by N for recall.
The script evaluates every combination of the listed input values, retains unrounded calculations and rounds displayed counts to whole assets and rates to a tenth of a percentage point. Row bounds are separate minima and maxima. S4 supplies the Wilson formula; S5 supplies exact binomial limits; S6 supplies the fixed-test design principles. Sample sizes and outcomes are Anvilfield audit cases. The script evaluates 243 parameter combinations. Each row reports its own minimum and maximum; the endpoints are marginal sensitivity bounds, not a jointly attainable low or high inventory or statistical confidence intervals. The central column is one complete case.
The downloaded qualitative inputs are CDC identification guidance updated in 2026 and the CDC equipment gallery dated 2024. The NIST interval page is undated and was accessed in 2026; its exact-limit reference is dated 2010, and the performance-threshold note is dated 2019. These sources define the evidence and statistical methods, without adding assumed commercial market totals.
In the procurement method, PorchRocket's Asset Search scopes custom searches and monitoring of visible assets, with each search quoted to its scope. Manufacturers, distributors and service companies should carry the same asset definition, dated evidence and acceptance rules from that scope into their own audit and sales handoff.
S1: Wong et al., automated cooling tower detection, author manuscript (2024)S4: NIST/SEMATECH, Confidence intervals (undated; accessed 2026)S5: NIST Dataplot, Exact Binomial Confidence Limits (2010)S6: Leber, Pibida and Enders, NIST Technical Note 2045 (2019)S2: CDC, Strategies for Identifying Cooling Towers (2026)S3: CDC, Pictures of Cooling Towers (2024)| Input | Values or rule | Assumption and scope |
|---|---|---|
| A1 | Normalize to 1,000 real, aerially visible cooling towers. | Fixed reference population and asset-level counting unit; hidden equipment is outside the visibility denominator. |
| A2 | Precision: 80.8%, 85.5%, 90.1%. Recall: 86.9%, 91.6%, 95.1%. | Independently vary published cohort values; use componentwise medians centrally. |
| A3 | Reject 70%, 85% or 95% of initial false positives. | Anvilfield review-effectiveness inputs. |
| A4 | Retain 97%, 99% or 99.5% of initial true detections. | Allow review losses; these rejected true assets remain missing. |
| A5 | Recover 25%, 50% or 75% of initial omissions. | Use a separate coverage search; accept additions only after independent corroboration. |
| A6 | Keep boundaries, observation date and identity rules fixed. | Match each asset once; add no unconfirmed recovery candidates. |
| A7 | Audit 100, 200 or 400 records at 95% observed correctness. | Independent random binary judgments, 95% Wilson intervals, no finite-population correction. |
| A8 | Preselect 59 checks and require zero errors. | Exact one-sided 95% error bound below 5%; no repeated peeking or clustered observations. |
FAQ
Is there one accuracy rate for aerial asset detection?
Use separate precision and recall results for the asset class, geography, imagery and matching rule being evaluated. The same detector changes performance across evaluation areas. A supplier score or a model confidence value does not establish the correctness and completeness of your territory inventory.
What is the difference between precision and recall?
Precision is the share of returned candidates that are correct. Recall is the share of real target assets found. Measure precision against independently checked candidates and recall against an independently established reference inventory that includes areas the machine did not flag.
Does human review remove all false positives?
Review catches look-alikes, duplicates and location errors when the evidence resolves them. Anvilfield estimates 23 residual false positives per 1,000 real visible towers, ranging from 5 to 68. Check residual errors independently and keep unresolved cases explicit.
Can a cleaner list still miss real assets?
Yes. Rejecting incorrect candidates improves precision while initially missed assets remain absent. Review also discards some genuine detections. Recovering omissions requires a separate search of eligible unflagged coverage and corroboration of the added assets.
How large should a quality sample be?
Set the claim, population, acceptable uncertainty and sampling design first. In the saved Anvilfield calculation, 190 correct records in 200 independent random checks gives a 95% Wilson interval of 91.0% to 97.3%. Use a design that accounts for clustering when several records share the same roof or campus.
Does a zero-error sample prove perfect detection?
No. A preselected sample of 59 independent checks with zero errors supports an error rate below 5% at 95% one-sided confidence for the sampled claim. A clean accepted-record sample does not establish that the search found every asset.
Can aerial review establish equipment condition or buying intent?
A reviewed sighting establishes a dated observation and its supported classification. Nameplates, equipment schedules, service records and an authorized survey establish the facts needed to price work. Confirm purchasing responsibility and the desired work directly with the appropriate contact.
What should a reviewed asset record contain?
Keep a stable identity, location, building and parcel association, capture date, visible evidence, proposed class, review disposition and correction history. Attach owner of record separately from operator and decision maker. Assign unresolved questions to a specific follow-up instead of filling them from appearance.
People also ask
Sources checked
- S1: Wong et al., automated cooling tower detection, author manuscript (2024) Checked 2026-11-05. Visible-tower precision and recall in the three evaluation cohorts; publication year 2024, original imagery collected in 2021, archive release 2025.
- S2: CDC, Strategies for Identifying Cooling Towers (2026) Checked 2026-11-05. Visibility, look-alike, image-age and human-review limitations; qualitative, no detection rate imported.
- S3: CDC, Pictures of Cooling Towers (2024) Checked 2026-11-05. Dated April 17, 2024 on the page; equipment examples for reviewer calibration; qualitative only.
- S4: NIST/SEMATECH, Confidence intervals (undated; accessed 2026) Checked 2026-11-05. Wilson interval formula, evaluated with Anvilfield sample assumptions.
- S5: NIST Dataplot, Exact Binomial Confidence Limits (2010) Checked 2026-11-05. Exact binomial inversion and one-sided limits; zero-error sample calculation.
- S6: Leber, Pibida and Enders, NIST Technical Note 2045 (2019) Checked 2026-11-05. Predeclared performance threshold, sample plan and acceptance rule; qualitative test-design basis.
- S7: PorchRocket, rooftop HVAC building lists (live 2026) Checked 2026-11-05. Subscribe & save tier prices $10 to $20 per building; delivery within 48 hours after payment; rooftop equipment, size and business filters.
- S8: PorchRocket, Asset Search (live 2026) Checked 2026-11-05. Custom searches and monitoring, reviewed visible-asset results, owner of record and scoped quotations; no dollar price imported.