The data
The starting point was the Pope County Assessor's parcel layer for the City of Russellville. Eleven thousand one hundred forty-two parcels, each with an assessed value, an owner, an address, and a polygon geometry.
The trouble was that the assessor's TAXAREA field fails in two directions. It was zero for 7,452 of those parcels, roughly two-thirds of the dataset. On many of the rest it was populated but implausibly small, because the GIS polygon captured only a subdivided piece of the actual lot. Lowe's is the cleanest example: polygon area reads 0.73 acres, TAXAREA reads 18, the truth is 18.
I rebuilt area from geometry directly, reprojecting each polygon to UTM Zone 15N and computing planar area in square feet, then converting to acres. VPA is assessed value divided by the larger of TAXAREA and the rebuilt polygon area. When the two disagree, the bigger number wins.
A residential parcel under 0.025 acres almost certainly isn't a real parcel. It's a fragment from how the county draws subdivision boundaries against irregular geometry. I flagged those as geometry artifacts and rendered them flat and gray. Two specific spikes that surfaced in early renders, one on Sky Ridge Road and one on South Jackson Street, are both caught by that rule.
The map
The map is one HTML file. The 3D layer is deck.gl extruding each parcel polygon to a height proportional to its VPA. The basemap is Carto Dark Matter via MapLibre. Color is a six-stop Urban3-style scale interpolated in LAB space, with a power-curve transform so the difference between a $700K parcel and a $4M parcel is visually distinct rather than compressed into one shade of orange.
Two design decisions worth naming. First, residential parcels show VPA and the label "Residential parcel" in the tooltip and nothing else. No address, no owner name. The map is public and there is no civic reason to surface that. Commercial parcels show the full record. Second, there is a true 2D choropleth mode in addition to 3D. A flat map and a height map answer different questions, and the toggle lets the viewer pick.
What the map shows
The downtown core, the 100-700 block of W Main where the small-lot commercial buildings cluster, averages $2.2 million per acre. Most of those parcels are under half an acre. That is what the city's most efficient land use pattern looks like.
St. Mary's Hospital hits $4.9 million per acre on a 5.87-acre footprint. It is not the city's top performer. A Simmons Bank branch at 800 N Arkansas leads the city at $9.7 million per acre on 0.18 acres. What makes St. Mary's notable is that it is the only large-footprint property in the city's top ten. Every other top performer fits on under 1.4 acres. Compactness is the variable.
Apartments outperform big-box retail on VPA. Spring Lake Apartments hits $548,000 per acre on essentially the same 17-acre footprint as Lowe's, which sits at $299,000. That is 83% higher. Single-family residential at $413,000 per acre median also beats Lowe's. Density doesn't have to mean tall to mean efficient.
Twenty-two percent of city land generates zero tax revenue. Government, churches, schools, cemeteries. That number isn't a complaint, it's a baseline. It tells you what the rest of the city has to carry.
The most actionable pattern is on West Main Street. About 15 acres of vacant and underutilized commercial parcels along that corridor are currently assessed at roughly $66,000 per acre. At the W Main commercial median of $663,000 per acre, those same 15 acres would carry about $9.9 million in assessed value. That is a $9 million opportunity on the city's best street, with no annexation required.
What this doesn't do
Assessed value lags market value, sometimes by years. VPA built on assessment isn't a real-time signal, it's a structural one. A parcel doesn't show up as undervalued the day after a sale, it shows up at the next reassessment cycle.
The map is also not a policy recommendation. It's a question generator. The right thing to ask of a low-VPA parcel on a busy corridor is what's preventing it from being something denser. That answer is usually in the code, not the dataset.