bshepp/residuals-fingerprints

RESIDUALS — LiDAR DEM residual fingerprints

39,716 residual images extracted by applying 593 distinct decomposition configurations × 25 upsampling methods to a single Fairfield County, Ohio LiDAR-derived Digital Elevation Model (1500×375 at 3.33 ft/px). Each row pairs a 256×256 PNG of the residual (rendered with the standard RdBu_r colormap, 99th-percentile symmetric clipping) with the algorithm and parameters that produced it, plus a 40-dim signature vector and pre-computed 2D/3D UMAP coordinates.

The companion source-code project is RESIDUALS; the rendered atlas / sweep videos / Blender flythrough that visualize this dataset are produced by the residuals-visuals project.

Why this dataset is interesting

• Algorithm classification benchmark: predict which decomposition algorithm produced a residual (24 family classes, or 593 fine-grained classes). A non-trivial visual task — many algorithm families produce visually similar outputs at certain parameter settings.
• Algorithm fingerprinting / signal-processing forensics: the included 40-dim signatures (radial-FFT power spectrum + statistical moments) cluster algorithm families clearly in UMAP space — useful as a reference for inverse-problem and provenance work.
• Same scene, all algorithms: every residual is a different mathematical lens on the same underlying terrain, making this an unusually clean substrate for studying what each algorithm preserves vs. destroys.
• Reproducible: source DEM hashes + RESIDUALS source code = full regeneration of the original 4.28 TB float64 outputs.

Schema

Column	Type	Description
filename	string	Original .npy filename in the source RESIDUALS exhaustive run
decomp_family	string	e.g. gaussian, wavelet_biorthogonal, morphological_rect, anisotropic_diffusion
decomp_params	string (JSON)	e.g. {"sigma": 10} or {"wavelet": "bior3.5", "level": 3}
upsamp_method	string	e.g. bicubic, lanczos, fft_zeropad, sinc_hamming
upsamp_params	string (JSON)	e.g. {"scale": 2, "kernel_size": 8}
scale	int32	Upsampling factor: 2, 3, 4, 8, or 16
category	string	Meta-category: classical, edge_preserving, wavelet, morphological, trend_removal, multiscale
signature	list[40]	32-bin radial-FFT log-power spectrum + 8 statistical moments (mean, std, skew, kurtosis, p50|abs|, p99|abs|, edge density, lag-1 autocorrelation)
umap_2d	list[2]	2D UMAP embedding of the signature
umap_3d	list[3]	3D UMAP embedding of the signature
image	binary (PNG)	256×256 PNG of the residual, RdBu_r colormap, 99th-pctile symmetric clip

Source data

• DEM: 1500×375 array, 3.33 ft/px resolution, derived from LiDAR tiles BS000600.las through BS000603.las covering ~1 mi² in Fairfield County, Ohio. Z range: 808.5–1034.4 ft. CRS: Ohio State Plane South (EPSG:3735).
• Residual crop: 640×640 region at scale=2 coordinates (row=1600, col=400), normalized to 256×256. Region was selected for high feature density (sweeping diagonal stream channel, branching drainage, V-shaped linear feature, distinct ridge/embankment patterns).
• 15 of the original 39,731 files are excluded due to division-by-zero edge cases in signature computation (all bicubic_16x_scale=16 outputs).

Splits

Single train split with 39,716 rows. The dataset is small enough that downstream users typically define their own splits (e.g. by decomp_family for held-out generalization, or random for IID evaluation).

Quick start

from datasets import load_dataset

ds = load_dataset("bshepp/residuals-fingerprints")
print(ds)
print(ds["train"][0]["decomp_family"], ds["train"][0]["decomp_params"])
ds["train"][0]["image"]  # PIL.Image.Image, 256x256

Citation

The canonical citation is the Zenodo archive (which contains the full data + visualizations + code snapshot):

@dataset{sheppard_residuals_2026,
  author    = {Sheppard, Brian},
  title     = {RESIDUALS: 39,731-residual exhaustive parameter sweep of
               decomposition × upsampling methods on a Fairfield County,
               Ohio LiDAR DEM},
  year      = 2026,
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19903273},
  url       = {https://doi.org/10.5281/zenodo.19903273}
}

Or this Hugging Face dataset (rendered subset for ML use):

@dataset{sheppard_residuals_fingerprints_2026,
  author    = {Sheppard, Brian},
  title     = {RESIDUALS — LiDAR DEM residual fingerprints (39,716 rows)},
  year      = 2026,
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/bshepp/residuals-fingerprints}
}

Project page: https://residuals.briansheppard.com · Source code: https://github.com/bshepp/residuals-visuals · Source pipeline: https://github.com/bshepp/RESIDUALS

License

Apache 2.0 — same as the source RESIDUALS project.

Limitations and ethics

• Single scene: all 39,716 samples come from the same DEM. Models trained here may not generalize to other terrains. Treat as a fingerprinting benchmark, not a generic remote-sensing pretraining set.
• Class imbalance: family counts range from 201 (polynomial) to 5,346 (wavelet). Use stratified splits or balanced sampling if classification accuracy matters.
• Known artifact: the leftmost ~30 columns of the source DEM exhibit upsampling-boundary effects. The crop window dodges this entirely.
• No PII / sensitive content: terrain residuals only. The source LiDAR is publicly-available Ohio state data.