davenquinn / cesium-martini

On-the-fly meshing of raster elevation tiles for the CesiumJS virtual globe
MIT License
81 stars 17 forks source link

Cesium-Martini

On-the-fly meshing of raster elevation tiles for the CesiumJS virtual globe

Himalayas

This package contains a Cesium TerrainProvider that uses right-triangular irregular networks (RTIN) pioneered by Mapbox's Martini to transform Terrain-RGB elevation tiles into quantized mesh terrain, for rendering in the CesiumJS digital globe. The module provides a general technique applicable to all raster imagery (although the Terrain-RGB format is near-ideal for streaming elevation data). Fixes for performance and better control of rendering quality are in progress.

This module was created to support our geologic map visualization work at Macrostrat and as a building block for future rich geoscience visualizations.

Installation

This package is listed on NPM as @macrostrat/cesium-martini. It can be installed using the command

npm install --save @macrostrat/cesium-martini

Cesium-Martini

Development

As of version 1.3.x, cesium-martini development is tested with the Yarn package manager. Your mileage with npm may vary.

Quick start

  1. Clone the repository
  2. Run git submodule update --init to fetch the martini submodule
  3. Install dependencies with yarn install
  4. Build the package with yarn run build

After cloning this repository, you can build the module (using Rollup) with yarn run build, or build and watch for changes with yarn run watch.

Example applications

Several example applications are available in the examples/ directory and runnable from the root project. The main example is built with Vite and others are built with Webpack v5. As well as showing how to use this module, these examples contain configuration for bundling Cesium in each packaging environment.

To run an example application, add MAPBOX_API_TOKEN=<your-mapbox-token> to a .env file. in the root of this repository, and then start the appropriate example:

Contributions in the form of bug reports and pull requests are welcome. These can be to add functionality (e.g. optional normal-map generation) or for performance. See list of known limitations below.

Motivation

The Cesium digital globe is a powerful platform for visualization of geospatial data in 3D. Cesium maintains a global elevation dataset as a prebuilt terrain mesh, which caches the computationally-intensive step of meshing height-field data into a triangle irregular network (TIN). Unfortunately, this quantized mesh format is relatively new, narrowly supported and tailored to Cesium itself. Going forward, supporting a TIN format for elevation datasets requires maintenance of significant single-purpose processing pipelines and storage resources.

Mapbox maintains a multiscale global elevation dataset in their clever terrain-RGB format, which bridges web standard file formats (PNG images) with traditional raster GIS formats for representing terrain. Rasters are the standard representation of elevation data across the geosciences, and many pipelines are available to create and modify raster images. Basically any elevation dataset can be easily rescaled to the Terrain-RGB format, but the jump from there to a "Quantized mesh" is more complicated.

Recently, the MARTINI project by Vladimir Agafonkin at Mapbox demonstrated an elegant algorithmic approach that sidesteps this issue. MARTINI meshes based on right-triangulated irregular networks (RTIN, Evans et al., 1998) and is far quicker than the traditional TIN generation techniques.

A speedy meshing algorithm allows this data-preparation step to be handled in the browser after elevation tiles are loaded. Integrating this toolchain into the Cesium digital globe enables the usage of Mapbox global data and other raster terrain layers (e.g. planetary and bathymetric data!), without adding overhead of TIN processing and storage.

Current limitations

Lack of support for overzooming

Cesium's implementations of the TerrainProvider interface are generally geared towards representing static terrain meshes. The RTIN algorithm used here can dynamically build meshes at a variety of error levels, and the input height field are data-dense and can represent extremely detailed meshes at a given zoom level. Currently, meshes are generated at levels of detail that undersample the available structure in a terrain tile — levels of detail are calibrated to what Cesium needs to render visually pleasing output at a given zoom level.

A smarter and more parsimonious solution would use much lower zoom levels for terrain than imagery, using the full resolution of the dataset in mesh construction. Done correctly, this could lead to an extremely data-efficient and adaptive terrain render, but this seems to run somewhat counter to how Cesium internally manages levels of detail. Ideally, someone familiar with the inner workings of Cesium would provide some guidance here.

Outstanding bugs and issues

Prior art and relevant examples

TODO

Pull requests for any and all of these priorities are appreciated!

Changelog

[1.3.0]: September 2023

Development environment

We reorganized the development environment and examples through a set of interrelated changes, for a more modern overall design.

[1.2.0]: November 2021

[1.1.3]: June 2021

[1.1.2]: May 2021

[1.1.0]: May 2021