Utility Vegetation Management Records for AI: Inspections, Prescriptions and Tree Work

This guide sits in our industry-specific operational data hub. It covers the records layer; aerial capture is covered in drone inspection imagery for utility and power line assets.

What a vegetation management record set actually contains

A usable vegetation dataset is a chain of linked events per span, not a single table. Most utilities run vegetation work in a dedicated work-management system or a GIS-centric field app, with contractor crews recording completions in the same or a parallel tool. Vendor and integrator overviews list maintenance records alongside imagery, LiDAR, GIS, weather, asset condition and terrain as model inputs [2], and imagery products report per-corridor distance to conductor and tree heights [3]. The records you want are the ones that say what an inspector saw, what was prescribed, and what crews actually did.

Expect these record types, in roughly this order of value for training:

• Inspection findings: patrol type (routine, enhanced, post-storm, hazard tree), date, inspector, span or circuit segment, tree count, species, diameter at breast height, height, lean, condition (dead, dying, diseased, structurally defective), and measured or estimated clearance to conductor.
• Prescriptions: the work unit assigned per tree or segment (trim, remove, mow, brush, herbicide, monitor), priority code, due date and the clearance target the trim must achieve.
• Work completion: crew, completion date, units completed, cleanup status, and any deviation from the prescription.
• QA and QC audits: sampled spans re-inspected after work, pass or fail, defect type (under-trim, missed tree, unauthorized removal).
• Constraints and refusals: property owner refusal, access denial, environmental or permit holds, and the follow-up action.
• Outcome events: vegetation-caused outages, momentary faults, wire-down events and ignitions, tied to a device and a location.

Inspection notes are often free text written in a truck. That makes them a strong source for extraction models, in the same way that inspection notes become image captions for vision-language work.

Which regulations shape the fields you will see

The regulatory regime a utility works under determines inspection cadence, clearance targets and what gets documented, so it also determines label meaning. On transmission, the NERC reliability standard FAC-003 (confirm the version in effect as of October 2026) requires owners to keep vegetation out of the Minimum Vegetation Clearance Distance (MVCD), to inspect applicable line miles on a recurring schedule and to complete identified work [6]. Transmission records under FAC-003 therefore tend to be well-structured, auditable and dense with clearance measurements, because they are built to survive a compliance audit.

Distribution records follow state rules, and California is the most documented case. CPUC General Order 95 Rule 35 sets radial clearances, with stricter values in the High Fire-Threat District, and investor-owned utilities layer enhanced vegetation programs on top through wildfire mitigation plans reviewed by the Office of Energy Infrastructure Safety [7]. Individual utilities also set local targets above the state minimum based on species growth, wind and snow load. Confirm the current rule text and each utility's plan directly; the values change through rulemakings.

Two consequences matter for a buyer. First, a "clearance compliant" flag means different things under FAC-003, GO 95 Rule 35 and a utility's own enhanced program, so you must capture the governing rule per record. Second, tree work practice is usually written to the ANSI A300 tree care standard [8], so prescription vocabularies (reduction, directional pruning, removal) are fairly consistent, but utilities still map them into their own work-unit codes.

How to build outcome labels from outages and ignitions

Outcome labels come from joining vegetation-caused outage and ignition events to the same span keys used by inspection and work records. Research on predicting vegetation-related distribution outages combines utility outage records with weather and vegetation management history, which is the pattern most commercial risk models follow [1]. The hard part is the join, not the model.

Plan for these failure modes:

• Cause-code drift: outage cause codes ("tree contact", "tree fell in ROW", "tree from outside ROW", "unknown") change between dispatch and final review. Use the final cause and keep the original.
• Device-level location: outages are often logged against the protective device that operated, not the span where the tree hit. You need an upstream device-to-span mapping, and the dates of switching changes.
• Survivorship in trim cycles: spans worked recently have fewer outages partly because they were worked. Without the work history, a model learns that recent trimming predicts low risk and nothing about why.
• Off-right-of-way trees: many vegetation outages come from trees outside the managed corridor. If inspections only record in-corridor trees, the label and the features describe different populations.
• Storm contamination: major event days inflate counts. Keep the event-day flag rather than deleting those rows, so you can test both ways.

Restoration-side fields such as crew dispatch, switching and customer minutes are covered in utility outage tickets and restoration logs; this page assumes you join those to vegetation records by device and time.

A request specification you can send to suppliers

A precise specification gets a faster and more honest answer from a data holder than a category name. The template below names the joins, the grain and the outcome window explicitly.

Illustrative example: invented to show structure; it does not describe an available dataset.

And an illustrative joined record, the shape a training row often takes after preparation:

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "finding_id": "VF-000123",
  "span_id": "SP-48812",
  "circuit_id": "C-1104",
  "inspection": {"type": "enhanced", "date": "2024-05-14", "species": "Quercus agrifolia",
                 "dbh_in": 22, "height_ft": 38, "condition": "structurally_defective",
                 "clearance_ft": 6.5, "note": "co-dominant stem, included bark, leaning toward phase B"},
  "governing_rule": "GO95_R35_HFTD",
  "prescription": {"work_unit": "REMOVE", "priority": "P2", "due": "2024-07-15"},
  "completion": {"date": "2024-06-30", "status": "completed", "deviation": null},
  "qa_audit": {"sampled": true, "result": "pass"},
  "constraint": {"type": null},
  "outcomes_24m": {"veg_outages": 0, "ignitions": 0, "major_event_days": 1}
}

Privacy and location: what to remove and what to keep

Location is the core feature in vegetation data, so de-identification has to remove people without destroying geography. Spans, poles and GIS geometry are utility assets, not personal data, but the records around them carry property owner names, phone numbers, parcel addresses, account numbers and refusal notes that can quote a customer. Those belong out of a training set, or replaced with consistent tokens so that repeat refusals at the same parcel still link. Our guide to de-identified data for AI covers methods and residual risk.

Practical rules for a buyer's specification:

• Keep span and pole coordinates at the resolution the model needs; ask whether parcel-level points can be snapped to the span.
• Require names, phone numbers, emails and account numbers removed or tokenized, with the method documented.
• Treat refusal and complaint notes as high-risk free text; ask for a reviewed sample before accepting the whole field.
• Ask whether contractor crew names are present; they are personal data too.

For general location data licensing patterns, see geospatial and location data.

Checking label quality before you train

Vegetation labels are noisy in predictable ways, and a quality check should measure the noise rather than assume it away. Audits of popular benchmarks found label error rates of at least 3.3% on average, enough to change model rankings [5]; vegetation cause codes and condition ratings are less standardized than those benchmarks. Use the supplier's own QA audit sample as the first estimate of completion-record accuracy, and compare inspector condition ratings on spans inspected twice within a short window to estimate rater agreement.

Ask for documentation in a Data Card style: upstream systems, collection methods, who entered each field and intended use [4]. Then run your own checks from our training data quality guide: duplicated findings across patrol types, prescriptions without completions, completions dated before inspections, and outage rows that fail the span join. For broader maintenance work orders outside vegetation, see maintenance work order datasets and inspection report data.

Rights questions specific to vegetation data

Rights in vegetation data are split across the utility, its contractors and its imagery vendors, so confirm who owns each layer before pricing anything. Contractor field apps may sit on the contractor's platform under its own terms. Imagery and LiDAR are frequently licensed to the utility for operational use only, and derived products such as tree-height rasters can carry the vendor's terms. Ask for the chain of rights per layer, and keep any imagery join out of scope unless that license is confirmed.

This is where a structured sourcing process helps. SourceX sources operational datasets from US companies on request, rather than holding them in stock, and manages the licensing process. Each dataset is rights-reviewed for ownership and consents before delivery under a license that defines records, uses, term and delivery. Buyers can describe the vegetation records they need without naming a utility; SourceX looks for businesses that hold that data, and every release is approved by the supplying company.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Request vegetation management data for AI

SourceX sources vegetation inspection, prescription and tree-work records from US companies on request, with personal details removed or replaced before delivery and nothing contracted until a supplier agrees. A request does not guarantee a match, but a specific description of spans, fields and outcome windows gives suppliers what they need to assess it. Describe your requirements at sourcex.si/buyers.