KDE density splat

Turn a weighted point cloud into a smooth Gaussian kernel-density surface on a grid.

What it means

A continuous picture of “how much point-stuff is here” — the windowed, artefact-free alternative to counting points in quadrats. Once the cloud is a grid, every image-front primitive (blur, hotspots, thresholding, wavelets) applies unchanged. This is the bridge between the two fronts.

GPU approach

The no-atomics render path: each point is drawn as a small instanced quad with a Gaussian footprint, additively blended into an r32float render target. The blend hardware accumulates overlaps — no atomic<f32> needed. Written in raw WebGPU so it ports to deck.gl / MDV / SpatialData.js as a custom layer.

Usage

import { splatDensityGpu } from 'intraspatial/gpu/spatial/splatDensity';

const field = await splatDensityGpu(xs, ys, {
  width: 256, height: 256, sigma: 8, // bandwidth in world units
  weights,                            // optional per-point weight
});
// field.data: Float32Array (row-major), field.bbox: world bounds

Composes with

• Separable convolution — further smooth or window the density.
• Getis-Ord hotspots — find where the density is significantly elevated.

Status: ✅ implemented and validated vs a CPU KDE golden · src/gpu/spatial/splatDensity.ts