dtmthinner.h File Reference

#include "map/mapextent.h"
#include <QPointF>
#include <QString>
#include <QVector>
#include <functional>

Include dependency graph for dtmthinner.h:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Classes
struct	mesh::DTMThinnerLimits
	Test-injectable resource limits for generatePoints(). More...
struct	mesh::DTMThinnerOptions
	Parameters for the normal-deviation terrain thinning. More...
class	mesh::DTMThinner
	Terrain-adaptive Steiner point selector from a DTM raster. More...

Namespaces
namespace	mesh

Typedefs
using	mesh::DTMProgressFn = std::function< bool(double)>
	Progress/cancel callback for DTMThinner::generatePoints().

Detailed Description

Author

Caleb Buahin caleb.nosp@m..bua.nosp@m.hin@g.nosp@m.mail.nosp@m..com

Date

2026

License\n GPL-3.0-or-later

Slice AU.5b — terrain-adaptive Steiner point generator.

Algorithm: iterative normal-deviation decimation on the full DTM grid.

Each pass: For every interior active vertex V:

1. Build the fan of triangles from V through its active 8-ring neighbours. All consecutive active-neighbour pairs form a face, including the closing "hole-bridging" pair — no gap check is applied. Downward face normals are flipped so all normals point into the upper half-space.
2. If fewer than 3 active ring neighbours remain, V is conservatively kept (score = 0) — too few neighbours to score reliably.
3. Compute the area-weighted vertex normal N_V from all fan faces.
4. score = min (or avg) of dot(N_V, face_normal_i) across all fan faces.
5. score >= normalDotThreshold → smooth here → MARK for removal. score < normalDotThreshold → terrain feature → KEEP. Batch-remove all marked vertices simultaneously. Hole retriangulation is implicit: the next pass builds the fan from still-active neighbours only, spanning across the gaps via the closing triangle. Only neighbours of removed vertices need rescoring (dirty-set optimisation).

Boundary pixels are NOT special-cased: the sampling grid is offset by half a step from the domain edge so every point is strictly interior, and the domain outline is carried by the PSLG segments instead. All grid points are therefore treated uniformly by the thinning algorithm.

Scaling notes (2026-07-31, banded thinning): • The DEM is read in horizontal strips sized to a fixed byte budget, so peak raster-buffer memory is independent of DEM and domain size. Interpolation is bit-identical to a single whole-bbox read. • Grid points outside the DEM footprint yield NaN and become inactive — they are never edge-clamped into a fabricated elevation. • The thinning GRID is processed in row BANDS bounded by a per-band working-set ceiling (~46 bytes per grid point), so total grid size is unlimited — a 15 GB DEM at native spacing streams through band by band. • Band seams are exact: the decimation is a radius-1 stencil per pass, so each band carries a halo of passes rows per side and emits only its core rows — bit-identical to an untiled run for any finite pass count ≤ kMaxThinningHalo. "(unlimited)" passes are capped at kMaxThinningHalo per band in multi-band mode (deterministic; single-band grids keep true run-to-convergence semantics). • Total retained points are capped (kMaxRetainedPointsDefault); exceeding the cap fails with an actionable message in errorMsg() rather than letting the downstream PSLG/Triangle stage exhaust memory.

Threshold guide (dot product = cos θ): 0.99 → keep bends > ~8° (fine detail) 0.95 → keep bends > ~18° (default — channels, levees, ridges) 0.90 → keep bends > ~26° (coarse — prominent breaks only)