Kyle Walker

Client-side #GIS tools are very effective for exploring patterns in dense datasets.

Shown here: click a Texas well, create an AOI from the approximate boundary of other wells in its lease / unit.

Stack: R, MapLibre/mapgl, Turf.js.

Looks great!

Finishing a lab right now for GIS students where they use Experience Builder and vibe code apps, then compare the two

Banging my head against the keyboard using Experience Builder

ChatGPT one-shotted a much better app with the JavaScript SDK

(which I can't write myself by hand)

We're at a weird inflection point where LLMs are making it easier to build dashboards / data apps with code than with WYSIWYG / drag-and-drop editors

The advantage to WYSIWYG was always the gentle learning curve compared to more customizable code... that advantage is evaporating

I'm a loyal Claude Opus / Claude Code user, but I have to say...

They are right about Codex 5.3

My most popular map ever, Educational Attainment in America, gets a fresh look for 2026:

- Updated to the 2024 ACS
- MapLibre via R's mapgl with built-in interactivity
- OpenFreeMap basemap, Cloudflare-hosted PMTiles (blazing fast!)

Explore your city: personal.tcu.edu/kylewalker/m...

It's easy to feel overwhelmed by the current pace of change in technology.

A useful counterpoint:

There are STILL organizations using ArcMap for their GIS even though ArcGIS Pro has been out for over a decade.

Technology moves fast; capacity for adoption, not always.

Showed my Advanced GIS students Claude Code this week.

We built this map of code compliance violations in Fort Worth; heatmap giving way to points on zoom.

Used the ArcGIS SDK for JavaScript.

Took 10 minutes.

I believe in Jevons Paradox.

But we need to be teaching skills for 2026, not for 2016.

West Texas energy infrastructure identified from satellite imagery using zero-shot image segmentation.

Shown:

Well pads from simple text prompts;

Draw a box around a wind turbine, find the others in the view;

Click a solar farm, segment the area.

Half of the United States' population lives in these 144 counties.

Explore: walkerke-county-popu...

In Feb 2024, I asked ChatGPT:

"What do income levels by neighborhood look like in Dallas, Texas? Make me a map of the results."

The result: the first nonsensical image.

Today I asked Claude that question - it made an accurate MapLibre map.

Wonder where we'll be in 2028...

Quality-of-life improvements to the #rstats {geosam} package:

- Interactive swatches let you customize extraction color in the explorer;
- The explorer sidebar is now collapsible;
- The package installs the stable version of HuggingFace Transformers

walker-data.com/geosam

walker-data.com/geosam

Distributed energy is harder to map than centralized infrastructure, but cutting-edge AI tools can help.

Shown: identifying rooftop solar panels in a suburban neighborhood with Mapbox imagery, Meta's SAM3.

Useful for utilities, grid planners, and solar prospecting.

Mapped every pipeline in Texas: 368,000 miles of infrastructure at your fingertips.

For energy operators, midstream companies, and data center site selectors, understanding pipeline proximity matters.

Open data + modern GIS makes it accessible.

Learn more:

walker-data.com/tidycensus

walker-data.com/mapgl

library(tidycensus)
library(mapgl)

get_acs(
geography = "tract",
variables = "B19013_001",
state = "TX",
geometry = TRUE,
year = 2024
) |>
maplibre_view(
column = "estimate",
interactive_legend = TRUE
)

It's 2024 5-year American Community Survey release day!

The ACS is the best resource for small-area analysis of US demographics.

In R, use tidycensus and mapgl to dive into the data right now.

Copy-paste the code below for an interactive income map with a legend filter:

Interactive data visualization dashboard showing Palmer Penguins dataset. Left sidebar contains collapsible controls for selecting X axis (Bill length), Y axis (Bill depth), and color grouping (Species) variables using dropdown menus. Main panel displays a scatter plot with points colored by penguin species (Adelie, Chinstrap, Gentoo), built with Esri's Calcite Design System components in R Shiny.

{calcite} looking good 💅

🗺️ Esri's Calcite Design System in {shiny}? Yes, please!!!

Full-stack spatial in R keeps getting better.

github.com/r-arcgis/cal...
#rstats

The big story coming out of yesterday's 2025 Census Population Estimates release: international migration declined in EVERY state in the past year.

The plot: international migration rates since 2021 (each year covers the previous 12 months).

Biggest drops: NY, NJ, CO, TX, UT

The 2025 @uscensusbureau Population Estimates were released today. Some takeaways:

- Rapid growth in the Carolinas, Idaho, Texas;
- 17 states + PR had more deaths than births;
- Florida is #1 for international migration;
- SC ranks #1 for domestic migration rate; NY ranks last

I merged a PR for mirai today (fixing an esoteric bug), that came with a performance boost that I'd never have thought existed. That means that on my laptop, the default (with dispatcher) round-trip performance now dips into sub-100 microseconds territory!! Get it now: `pak::pak("r-lib/mirai")`

Nearly 18 million California jobs in 2023, visualized by 2-digit NAICS category with block-level precision.

Open data and modern geospatial technology make powerful labor market analytics possible.

Identifying possible empty lots in Fort Worth's South Main district with Meta's Segment Anything Model 3.

Draw an "exemplar" prompt, find similar areas in the satellite image. Useful for urban redevelopment, real estate, and more.

Try out {geosam}: walker-data.com/geosam

Social media: "if you're not using {Ralph Wiggum|Clawdbot|next week's new framework} you're ngmi"

Now you're endlessly trying to optimize your workflows instead of doing the actual work you're avoiding.

Agents will get better and better but they don't substitute for agency.

spopt for R implements the Max-P-compact regions algorithm. Draw contiguous regions that optimize both for attribute similarity and compactness.

Shown here: regular Max-P with queen weights (left), Max-P-compact with rook weights (right)

Learn more: walker-data.com/spopt-r/arti...

awesome! lonboard is quite impressive. My demos use PMTiles but there's more setup involved for that

if you need it now but can't install from source, try R-Universe!

walkerke.r-universe.dev/mapgl

Southlake, Texas

definitely! it is getting better, I'll admit. Google / Gemini is probably the best at it.

LLMs both enable and impede innovation in open source software.

They help developers bring previously impossible ideas to life.

But they also recommend older solutions to users given their training data.

So even if a new innovation is better, users might not know about it.