Files
rfcp/compass_artifact_wf-557af7f1-f9ba-4e64-ac94-f28e56ef6cd8_text_markdown.md
2026-02-06 22:17:24 +02:00

118 lines
12 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Smooth RF coverage maps in WebGL: a production playbook
**Replace one line of shader code and your grid squares vanish.** The single highest-impact fix for pixelated RF coverage overlays is swapping hardware bilinear texture sampling for a **Catmull-Rom bicubic fragment shader**—9 texture fetches instead of 1, near-zero performance cost, and mathematically correct C1-continuous interpolation that passes through your actual RSRP values. Professional RF tools like Atoll and CloudRF sidestep the problem entirely by computing at every grid cell, but when you're rendering a sparse client-side grid (1,9756,675 points), shader-based interpolation is the correct approach. This report covers exactly how to implement it, what the industry does, and which open-source libraries can accelerate the work.
## How professional RF tools avoid the problem you're solving
Professional RF planning tools don't interpolate sparse data—they brute-force compute signal values at every pixel. **CloudRF** runs its SLEIPNIR propagation engine server-side at user-specified resolution (down to 1 m with LiDAR), outputs GeoTIFF or pre-colored PNG, and the client simply overlays the image via `L.imageOverlay` in Leaflet or drapes it on CesiumJS for 3D. There is no client-side interpolation. **Atoll** (Forsk) computes predictions at **550 m grid resolution** within its desktop GIS—smoothness comes from grid density, not post-processing. **SPLAT!** does the same with Longley-Rice/ITM, outputting PPM rasters where each pixel maps 1:1 to a DEM grid cell; its `-sc` flag adds smooth color gradients but no spatial interpolation.
The pattern is universal: propagation model → dense raster → colorize → overlay. Crowdsourced coverage services like **Ookla/Speedtest** (which uses Mapbox GL JS with WebGL rendering) and **OpenSignal** follow a different path: they aggregate sparse measurements into spatial bins, apply kernel density estimation or IDW smoothing server-side, then serve the result as raster tiles. The industry-standard interchange format is **Cloud-Optimized GeoTIFF (COG)** with bilinear or cubic resampling via GDAL at tile-generation time.
Your situation is fundamentally different from both. You have a moderate-density regular grid arriving from the backend, and you need the client to render it smoothly at arbitrary zoom. This makes shader-based interpolation the right tool—not denser computation, not server-side tiling.
## Catmull-Rom in a fragment shader: the production solution
The core insight is that your current pipeline—upload grid as float texture, sample with `GL_LINEAR`, apply colormap—is already 90% correct. Hardware bilinear interpolation produces C0 continuity (values match at grid edges, but derivatives don't), causing visible seams. **Catmull-Rom spline interpolation** provides C1 continuity (smooth first derivatives) while still passing through your exact data values, unlike B-spline bicubic which smooths/blurs peaks.
The 9-tap Catmull-Rom implementation, widely used in production (ported from TheRealMJP's gist with 108 GitHub stars), replaces your `texture2D()` call:
```glsl
vec4 SampleTextureCatmullRom(sampler2D tex, vec2 uv, vec2 texSize) {
vec2 samplePos = uv * texSize;
vec2 texPos1 = floor(samplePos - 0.5) + 0.5;
vec2 f = samplePos - texPos1;
vec2 w0 = f * (-0.5 + f * (1.0 - 0.5 * f));
vec2 w1 = 1.0 + f * f * (-2.5 + 1.5 * f);
vec2 w2 = f * (0.5 + f * (2.0 - 1.5 * f));
vec2 w3 = f * f * (-0.5 + 0.5 * f);
vec2 w12 = w1 + w2;
vec2 offset12 = w2 / (w1 + w2);
vec2 texPos0 = (texPos1 - 1.0) / texSize;
vec2 texPos3 = (texPos1 + 2.0) / texSize;
vec2 texPos12 = (texPos1 + offset12) / texSize;
vec4 result = vec4(0.0);
result += texture2D(tex, vec2(texPos0.x, texPos0.y)) * w0.x * w0.y;
result += texture2D(tex, vec2(texPos12.x, texPos0.y)) * w12.x * w0.y;
result += texture2D(tex, vec2(texPos3.x, texPos0.y)) * w3.x * w0.y;
result += texture2D(tex, vec2(texPos0.x, texPos12.y)) * w0.x * w12.y;
result += texture2D(tex, vec2(texPos12.x, texPos12.y)) * w12.x * w12.y;
result += texture2D(tex, vec2(texPos3.x, texPos12.y)) * w3.x * w12.y;
result += texture2D(tex, vec2(texPos0.x, texPos3.y)) * w0.x * w3.y;
result += texture2D(tex, vec2(texPos12.x, texPos3.y)) * w12.x * w3.y;
result += texture2D(tex, vec2(texPos3.x, texPos3.y)) * w3.x * w3.y;
return result;
}
```
**Performance is essentially free.** Benchmarks on a GTX 980 show 9-tap Catmull-Rom at 1920×1080 costs **~0.32 ms** versus ~0.30 ms for single-tap bilinear. Your ~80×85 texel coverage texture is trivial. The critical rule: **interpolate raw scalar RSRP values first, then apply the colormap**. If you interpolate after colorization, you get color-space artifacts (muddy intermediate colors between discrete bands).
For the absolute quickest improvement with zero extra texture fetches, Inigo Quilez's **smoothstep coordinate remapping trick** eliminates grid edges with a single line change:
```glsl
vec4 textureSmooth(sampler2D tex, vec2 uv, vec2 texSize) {
vec2 p = uv * texSize + 0.5;
vec2 i = floor(p);
vec2 f = p - i;
f = f * f * f * (f * (f * 6.0 - 15.0) + 10.0); // quintic hermite
return texture2D(tex, (i + f - 0.5) / texSize);
}
```
This gives C2 continuity with a single texture read, though it introduces slight positional bias. For a quick visual test before implementing full Catmull-Rom, it's ideal.
## When to use IDW instead, and how the GPU handles it
Catmull-Rom works because your data sits on a **regular grid**. If your backend ever returns irregular point distributions (e.g., drive-test measurements, crowdsourced data), **Inverse Distance Weighting (IDW)** on the GPU becomes the right approach. The production technique uses WebGL's additive blending to parallelize across data points rather than looping in a shader:
For each data point, render a full-screen quad where the fragment shader computes `w = 1/dist^p` and outputs `(w × value, w, 0, 1)` into a framebuffer with `gl.blendFunc(gl.ONE, gl.ONE)`. After N passes (one per point), a final shader reads the accumulated texture and computes `interpolated_value = R_channel / G_channel`. This avoids the O(N) per-pixel loop that would choke a fragment shader at 7,000 points.
The open-source **`mapbox-gl-interpolate-heatmap`** library implements exactly this pattern, with a `framebufferFactor` parameter (typically 0.30.5) that renders the IDW computation at reduced resolution for performance, then upscales for display. The **`temperature-map-gl`** library from ham-systems provides the same approach in a minimal ~3 KB package.
**For your regular-grid case, IDW is strictly worse than Catmull-Rom**: it's slower (N draw calls vs. 9 texture fetches), creates bull's-eye artifacts around data points, and doesn't exploit grid structure. Reserve it for irregular data.
## Mapping library internals and the tiled raster alternative
If you want to move beyond a custom WebGL overlay, understanding how major mapping libraries handle this problem reveals useful architectural patterns:
**Mapbox GL JS** exposes `raster-resampling: 'linear'` (bilinear) or `'nearest'` for raster tile layers—standard GPU texture filtering, same limitation you're hitting. Its built-in heatmap layer uses Gaussian kernel density estimation with additive blending into an offscreen half-float texture, designed for point density visualization, **not scalar interpolation**. However, Mapbox v3's `raster-array` source with `raster-color` expressions enables client-side colorization of raw data tiles, which is directly applicable. **deck.gl** offers `BitmapLayer` with configurable `textureParameters` (set `magFilter: 'linear'`) and supports full custom shader injection via `fs:DECKGL_FILTER_COLOR` hooks, but its `HeatmapLayer` is again KDE-based, not interpolation-based.
For a **tiled architecture** (useful if your coverage areas grow large), the proven pipeline is:
- Backend computes RSRP grid → stores as GeoTIFF
- GDAL resamples with `cubicspline` or `lanczos`: `gdalwarp -r cubicspline -tr 10 10 input.tif upsampled.tif`
- `gdal2tiles.py` generates XYZ tile pyramid
- Client displays via `L.tileLayer` with standard bilinear filtering
The **IHME `leaflet.tilelayer.glcolorscale`** library offers a sophisticated variant: encode 32-bit float values into PNG RGBA channels server-side, decode in a WebGL fragment shader client-side, and apply dynamic color scales without re-tiling. Its companion `leaflet.tilelayer.gloperations` adds GPU-based convolution smoothing. This pattern preserves raw values for pixel queries while enabling dynamic color ramp changes.
## Handling boundaries, gaps, and color mapping
Three practical concerns beyond core interpolation deserve specific solutions. For **coverage boundaries**, apply `smoothstep` fading based on signal strength or a validity mask:
```glsl
float signalStrength = SampleTextureCatmullRom(u_data, uv, texSize).r;
float boundaryAlpha = smoothstep(-115.0, -105.0, signalStrength); // fade near noise floor
gl_FragColor = vec4(colormap(t), boundaryAlpha * u_opacity);
```
For **missing grid cells**, store a validity mask as a second texture channel (or use a sentinel value like -9999). Interpolate both the value and mask with Catmull-Rom; the interpolated mask naturally creates smooth alpha transitions at data boundaries without hard edges.
For **color mapping**, use a **1D texture lookup** rather than branching `if/else` chains in GLSL. Upload your RSRP→color ramp as a 256×1 RGBA texture, normalize your interpolated value to [0,1], and sample: `vec3 color = texture2D(u_colormap, vec2(t, 0.5)).rgb`. This is faster than arithmetic color functions and trivially supports any color ramp. The `glsl-colormap` package provides standard scientific palettes (viridis, jet, etc.) as pure GLSL functions if you prefer avoiding the extra texture.
## Recommended implementation path
The optimal architecture for your Electron + React + Leaflet stack, given 1,9756,675 grid points:
**Immediate fix (30 minutes):** Replace `texture2D(u_data, uv)` with the smoothstep trick in your existing fragment shader. This eliminates visible grid squares with zero performance cost and zero architectural changes.
**Production implementation (12 days):** Implement the 9-tap Catmull-Rom `SampleTextureCatmullRom()` function. Pack your RSRP grid into a float texture (R channel = RSRP value, G channel = validity mask). Apply Catmull-Rom to both channels. Use a 1D colormap texture for color mapping. Add smoothstep boundary fading. This produces results visually indistinguishable from CloudRF-quality coverage maps.
**If you outgrow the single-texture approach** (very large coverage areas, multiple overlapping cells): transition to the float-encoded tile pipeline using `leaflet.tilelayer.glcolorscale` or generate server-side tiles with GDAL cubic resampling. The tile pyramid handles LOD automatically and scales to arbitrarily large coverage areas.
## Conclusion
The gap between your current output and professional RF coverage visualization isn't architectural—it's a single shader function. Professional tools achieve smoothness through brute-force grid density (computing at every cell), but shader-based Catmull-Rom interpolation produces equivalent visual quality from sparse grids at negligible GPU cost. The 9-tap implementation requires **9 texture fetches** versus your current 1, adds C1 continuity that eliminates all visible grid edges, and preserves exact RSRP values at grid points—unlike Gaussian blur or B-spline smoothing, which distort the data. For truly irregular point data, GPU-accelerated IDW via additive blending is proven in production libraries like `mapbox-gl-interpolate-heatmap`. The critical implementation principle: always interpolate raw scalar values first, colorize second.