January 1, 2026

Ridge Rumble in Comment Canyon

Show HN: Feature detection exploration in Lidar DEMs via differential decomp

Dev drops a map tool to find hidden terrain, and sparks a nerdy filter war

TLDR: An open-source tool tests thousands of filter combos to reveal hidden features in elevation maps, complete with a cheeky “uncertainty of uncertainty” chart. The community split between loving its transparency and demanding neural networks, while debating whether hillshade is real truth—and laughing about laptops melting.

A new open-source drop called RESIDUALS hit Hacker News and the comments went volcanic. It uses laser-made elevation maps (LiDAR) to find hidden features like ridges and roads by mixing and matching filters, then comparing results to pull out what each combo reveals. The kicker? A cheeky “meta-divergence” chart—basically the uncertainty of uncertainty—which the crowd immediately turned into a meme.

Strong opinions flew fast. One camp cheered the transparent, reproducible grid of 39,731 filter-and-zoom combos, calling it “science, not vibes.” Another camp yelled “just use a neural network!”, accusing the project of reinventing image processing with fancy names. Then came the ground truth brawl: the tool compares to hillshade (a pretty shadowed map), and purists argued hillshade is vibes, not truth. Performance drama bubbled up too: “Running the full sweep made my laptop sound like a drone,” joked one user, while others formed micro-fandoms around Lanczos vs B‑Spline like it was the World Cup of filters.

Still, plenty of love: detailed docs, hashes for every run, and real-world uses—terrain analysis, infrastructure detection, change over time—won respect. The funniest moment? The “Known Limitations” section cutting off mid-sentence right after “Edge artifacts,” prompting quips like “even the limitations have edge cases.” Classic HN energy.

Key Points

  • RESIDUALS is an open-source framework that detects features in DEMs by combining decomposition and upsampling methods, then differencing outputs to isolate features.
  • A five-stage hierarchy (levels 0–4) compares ground truth, residuals, residual-vs-ground-truth, divergence, and meta-divergence to assess method behavior.
  • The framework includes numerous decomposition methods (Gaussian, bilateral, wavelet, morphological, top-hat, polynomial, etc.) and upsampling methods (bicubic, Lanczos, B-spline, FFT, and more).
  • An exhaustive runner evaluates 39,731 parameter combinations and generates documentation with statistics, hashes, and visualizations arranged in a grid.
  • Use cases include terrain, infrastructure, natural feature analysis, change detection, and quality assessment, with a noted limitation of edge artifacts in morphological methods.

Hottest takes

“Just throw a CNN at it” — ml_dude
“Hillshade isn’t ground truth, it’s vibes” — topo_nerd
“39,731 combos? My laptop started sounding like a drone” — fan_of_fans
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.