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

Summary

RESIDUALS: Multi-Method Differential Feature Detection A framework for feature detection in Digital Elevation Models using systematic decomposition and differential analysis. What Is This? RESIDUALS systematically tests combinations of signal decomposition and upsampling methods to identify which combinations best reveal features in elevation data. The core insight: Different method combinations have characteristic behaviors that selectively preserve or eliminate different feature types. By computing differentials between outputs, we create feature-specific extraction filters. The 4-Level Differential Hierarchy Level What It Shows Column 0 Ground truth (hillshade) DEM 1 Decomposition residuals bicubic, lanczos, bspline, fft 2 Residual vs ground truth Δbic, Δlan, Δbsp, Δfft 3 Divergence across methods Div 4 Meta-divergence (uncertainty of uncertainty) ΔDiv Quick Start # Clone and install git clone https://github.com/bshepp/RESIDUALS.git cd RESIDUALS pip install -r requirements.txt # Run with included test DEM python run_experiment.py # Or generate your own from LiDAR python generate_test_dem.py --lidar-dir /path/to/las/files --grid-rows 4 --grid-cols 4 python run_experiment.py --dem data/test_dems/your_dem.npy Decomposition Methods Method How It Works Good For Gaussian Local blur → subtract Smooth gradients, valleys Bilateral Edge-preserving blur Features with sharp boundaries Wavelet Multi-scale frequency separation Scale-specific features Morphological Shape-based opening/closing Peaks, ridges, depressions Top-Hat Small feature extraction Small isolated features Polynomial Global surface fitting Regional-scale variations Extended methods include: anisotropic Gaussian, median, DoG, LoG, guided filter, anisotropic diffusion, rolling ball, and multiple structuring element shapes. Upsampling Methods Method Characteristics Bicubic Balanced baseline (order 3) Lanczos Sharper edges, some ringing B-Spline Smoother (order 2) FFT Zero-pad Band-limited, Gibbs ringing at edges...