How to digitize as-built drawings into GIS: from scanned sheets to usable layers
Most utilities and public works departments know where the record of their pipes is: in a plan room, a filing cabinet or a folder of scanned PDFs. A scan keeps the drawing safe, but nobody can query it. You cannot ask a PDF which 18 inch lines were installed under a given street, select every inlet in a catchment, or hand it to a drainage model.
Getting from a scan to a GIS layer has usually meant someone tracing each sheet by hand. This guide covers the whole job in plain terms: what has to happen to a drawing before it counts as GIS data, where the common errors come from, how to check the result, and where AI now does the tracing and retyping for you.
Why scanned as-builts need to become GIS data
A scanned sheet is a picture of information. It has no records, no fields and no location that software can read. That gap shows up in everyday work:
- Permit mapping. Municipal stormwater permits commonly call for a map of the system. Massachusetts' small MS4 permit, for example, requires permittees to map outfalls, the drainage area to each outfall and the associated storm drain infrastructure. Requirements differ by state and permit, so check yours.
- Asset management. FHWA's digital as-builts program treats as-built data as part of the handover into asset management, and lists locating assets geospatially and moving asset data into GIS among the essential uses.
- Modeling. A drainage model needs pipes that connect, with sizes and flow line elevations. Those details are often on the sheet and nowhere else.
- Design and locates. Engineers want to see what is already in the ground before they draw a new alignment, without pulling sheets one project at a time.
FHWA's guidance is aimed mainly at new projects that deliver digital data from the start. The harder problem for most owners is the backlog: decades of drawings that already exist only as paper or scans.
As-built, record drawing, GIS layer: what you are converting
The words get used loosely, and the difference matters when you decide how far to trust the output. Definitions vary by agency, but this is the common usage:
| Term | What it usually means | What that means for GIS |
|---|---|---|
| Design plans | What was intended to be built | May not match the ground. Use only if nothing later exists |
| As-built or redline | Plans marked up by the contractor or inspector to show what changed | Closer to reality. Handwritten changes can be easy to miss |
| Record drawing | The final set prepared for the owner's records, usually incorporating those changes | Usually the best source. Check the date and any revision block |
| GIS layer | Features with geometry, attributes and a coordinate system | The goal: every pipe and structure is a queryable record |
If several versions of a project exist, digitize the latest and record which sheet each feature came from. That one source field saves a lot of argument later.
The traditional way, and why it stalls
Done by hand, the job runs like this in ArcGIS, QGIS or CAD:
- 1Scan the sheet, or find the existing scan.
- 2Georeference the image by picking matching points on the sheet and on a base map.
- 3Trace every pipe, structure and fitting as a new feature.
- 4Type the attributes from the callouts into each record: size, material, station, flow line.
- 5Check the work, then repeat for the next sheet.
None of it is conceptually hard. It is slow, and it is a poor use of a skilled technician. On a backlog of thousands of sheets it tends to lose out to every urgent project, which is why so many digitization efforts finish a pilot area and stop.
Step 1: identify the features and build a legend
First decide what you want out of each sheet. Write down the feature classes before anyone starts, because a clear list drives everything after it: water lines, storm lines, valves, hydrants, inlets, manholes, flow arrows and so on. Then match them to your existing GIS schema if you have one, so the new records land in the right place.
A drawing's own symbols are not universal. Line styles and symbols differ between engineers, agencies and decades, so a legend that works on one project may not work on the next. Sample sheets from different eras before you commit to a process.
NeuraCities works through the scan one feature class at a time. Each class is named, counted and colored, and the tool builds a legend for the sheet. Here is a real example: a single 2005 storm drain record drawing from North Texas, used on the Digitization product page.
Show the data
| Water lines | 20 |
| Valves | 15 |
| Storm lines | 14 |
| Inlets | 7 |
| Hydrants | 7 |
| Flow arrows | 6 |
Step 2: read the callouts and attach them to the right feature
Geometry without attributes is only a slightly better picture. The value of a record drawing sits in its text: station, pipe size and material, fitting type, flow line elevation and coordinates.
Plain OCR gets you the characters but not the meaning. It will not tell you that a string of numbers is a flow line, or which of three nearby pipes a leader line points to. The extraction step has to do three things: read the text, understand what the callout means, and attribute it to the correct feature. Here are four of the callouts read from the example sheet:
| Station | Description | Pipe | Flow line (ft) |
|---|---|---|---|
| 0+00 SD SP-1 | Connect to existing inlet A-9 | 18 in HDPE | 586.56 |
| 0+37.38 SD SP-1 | 10 ft curb inlet SP-1, 3 ft depth | 18 in HDPE | 587.25 |
| 0+69.76 SD SP-2 | Install 30 degree bend | 18 in HDPE | 585.02 |
| 1+04.96 SD SP-3A | Install 60 degree wye | 30 in HDPE | 582.11 |
Design the attribute table first
Keep a small, consistent set of fields rather than one free text column. At a minimum, consider feature class, size, material, station, flow line or invert, install or record date, and the source sheet. Keep the original text too, so a reviewer can compare what was read with what was parsed.
Step 3: georeference the drawing to the real world
Georeferencing moves the sheet from paper coordinates to ground coordinates. Everything on it then moves too. Good control comes from, in rough order of preference:
- 1Survey coordinates printed on the sheet: northings and eastings at stations, or marked control points.
- 2Project control or monuments that tie to a known coordinate system.
- 3Features visible both on the sheet and in current data: property corners, manholes, curb returns or road intersections. Useful, but only as good as the base layer.
Older drawings on a local project datum, with no coordinates at all, are the hard case. Expect to rely on the third option, and expect lower confidence.
How many control points?
A first-order (affine) transformation shifts, scales and rotates a sheet. Esri's guidance is that it needs at least three well-placed points. With exactly three, the fit is mathematically perfect and the error reads zero, which tells you nothing. Add more where you can, spread across the sheet rather than bunched in one corner, because spare points are what expose a bad one. Esri also warns that a low RMS figure does not by itself prove good registration.
The example sheet was placed with an affine fit from three control points marked on the drawing, so nobody had to visit the site. That suits a clean sheet carrying its own survey marks. Where more points are available, use them.
The three checks that prevent silent errors
| Check | Why it matters | What to do |
|---|---|---|
| Horizontal datum and zone | The same northing and easting land somewhere else under a different datum or State Plane zone | Look on the sheet's notes or the project survey. Do not assume the current city standard |
| Units | NIST retired the US survey foot after 2022 in favor of the international foot, but older drawings often use it | Confirm which foot applies. The two differ by about 2 parts per million |
| Vertical datum | Elevations on older sheets may be in NGVD 29, not NAVD 88 | Convert with NGS methods or a surveyor's published values before mixing with modern terrain |
Two parts per million sounds harmless, but it is multiplied by the coordinate value. On a State Plane northing like the 7,020,352 on the example sheet, reading survey feet as international feet moves the point by about 14 feet (our arithmetic). That is enough to put a pipe under the wrong curb, and nothing on the map will look wrong.
Step 4: check the result before anyone relies on it
Whether a technician or an AI did the tracing, check the output the same way. Sample more heavily on the first sheets from a new source or era.
- Overlay on imagery. Inlets should sit on curb lines and manholes in the street. Look for a consistent shift, which usually points to a datum or unit problem.
- Compare with visible structures. Spot check a handful of points against a field survey or GPS where accuracy matters.
- Test connectivity. Pipes should meet at structures. Gaps and dangling ends will break a network model.
- Check flow direction. Compare the flow arrows on the sheet with the flow lines. Water should run downhill along the pipe.
- Reconcile the counts. Compare features in the GIS against the callouts and the drawing's quantities, where it has them.
- Compare neighboring sheets. Lines should line up across match lines. A step at the boundary points to a georeferencing error on one of them.
Step 5: export in the formats people use
Decide who will use the data and work backward. Each format serves a different person:
| Format | Who uses it | Notes |
|---|---|---|
| GeoJSON | Web maps and developers | Open and easy to load. Works well one layer per feature class |
| Shapefile | GIS teams | Widely supported in desktop GIS |
| CSV | Engineers and asset managers | Attributes as a table, for spreadsheets or an asset system import |
| SWMM network | Drainage modelers | Structured for EPA SWMM, the free, open source drainage model |
Load the data into a staging layer first. Keep your authoritative GIS untouched until someone has reviewed it, and carry the source sheet field through so any record can be traced back to its drawing.
How NeuraCities Digitization does it
NeuraCities Digitization automates the middle of this workflow. You send scanned as-built or record drawings, and it runs four stages:
- 1Detect. The scan is worked through one feature class at a time. Features are identified, and each class is named, counted and colored in a legend.
- 2Extract. It reads text directly off the sheet, works out what each callout means and attributes it to the right feature, so values are not retyped by hand.
- 3Put on the map. It matches the surveyed points on the drawing, fits the transformation, georeferences every feature and places the sheet on a live map.
- 4Deliver. You get GeoJSON and Shapefile per feature class, with attributes as CSV alongside. Drainage sheets can also come back as a SWMM network you can run.
Those are the product's headline figures, set against manual digitization. Time and quality on your own drawings will depend on the sheets, so start with a small sample that includes your worst scans as well as your best.
A digitized drainage network also feeds directly into modeling. The Stormwater Engineering product runs storm studies on a city's real drainage network, so records that were stuck in a cabinet become an input to a flood study.
Planning a backlog project
- 1Start with a purpose. A permit map, a model or a capital project gives you a priority area and a clear definition of done.
- 2Inventory the drawings. List projects, dates, locations and scan status. Find out which sheets have coordinates and which do not.
- 3Set the schema and the checks before processing anything.
- 4Run a mixed sample: clean and messy, old and new. Measure what passes review, then adjust.
- 5Process by area, and check neighboring sheets where they meet.
- 6Assign an owner. Agree who keeps the data current as new as-builts arrive. Otherwise the backlog simply starts growing again.
What digitization cannot tell you
- Whether the drawing was right. A record shows what someone documented, which may differ from what is in the ground.
- What changed afterward. Repairs, abandonments and undocumented connections will not appear unless they have their own records.
- Current condition. Age and material can be read from a sheet. Condition needs inspection.
- Legal survey accuracy. A georeferenced drawing is not a boundary or utility survey, and it does not replace one.
Used with those limits in mind, a digitized record set is still a large step up from a folder of PDFs. It tells you where to look, and it makes the knowledge in those sheets available to everyone.
Frequently asked questions
How do I convert as-built drawings to GIS?
Scan the sheets, identify the feature classes, read the callouts into attributes, georeference the sheet using surveyed points or visible control, check the result, and export it as GIS layers. You can do this by hand in ArcGIS, QGIS or CAD, or use a tool such as NeuraCities Digitization to automate detection, extraction and georeferencing.
How do you georeference a scanned as-built drawing?
Match points on the sheet to known ground coordinates, such as stations with printed northings and eastings, survey control, or features visible in current data. A first-order fit needs at least three points. Use more, spread across the sheet, and confirm the datum, units and zone first.
What is the difference between an as-built and a record drawing?
Usage varies by agency, but an as-built is typically a marked-up plan showing what changed during construction, and a record drawing is the final set prepared for the owner, usually incorporating those changes. For GIS, use the latest version available and record which sheet each feature came from.
Can AI digitize engineering drawings?
Yes. NeuraCities detects each feature class on a scanned drawing, reads the callouts and attributes them to the right features, georeferences the sheet and exports GIS files. You should still review the output, especially numbers such as flow lines and pipe sizes.
How long does it take to digitize an as-built drawing?
NeuraCities' product figure is about 30 seconds per drawing, compared with a day of redrawing by hand. Actual results depend on the drawings, and reviewing the output takes additional time.
How accurate is digitized as-built data?
It depends on the source drawing, the scan and the control points, and no single figure applies to every sheet. Digitized records still reflect only what the drawing documented, so treat them as records rather than field-verified locations and spot check against a survey where it matters.
Can as-built drawings be turned into a SWMM model?
NeuraCities can return drainage sheets as a SWMM network you can run. Review connectivity, sizes and flow lines before relying on it, and check that the network ties into the rest of your system.
What file formats does NeuraCities Digitization export?
GeoJSON and Shapefile per feature class, with attributes as CSV alongside, plus a SWMM network from drainage sheets.
More guides on digitizing drawings
Still got drawings in a filing cabinet? See a real storm drain sheet become mapped, queryable GIS layers.
See Digitization