Latest revision as of 12:38, 22 July 2012

[edit] Natural Earth

[edit] Introduction

Marble currently uses the very old and outdated MWDBII dataset for its vector base map such as national borders and coastlines and we really need to replace it with more up-to-date data. However, MWDBII has two key advantages, it is very compact in size enabling Marble to ship it by default, and the individual nodes have a zoom level value which speeds up drawing.

The current vector layer also has the disadvantage of not being able to be manipulated either programtiacally or by the user. This prevents it from being used for such things as KGeography or other educational uses where you would want to select and manipulate a geographic entity.

Improving the vector base maps would thus consist of two closely related parts:

Improving the vector drawing code to allow interaction with the vectors, and improved performance to allow more detailed vectors to be drawn.
Importing a new base layer dataset.

[edit] Using NaturalEarth

The Natural Earth data set is a "public domain map dataset available at 1:10m, 1:50m, and 1:110 million scales. Featuring tightly integrated vector and raster data, with Natural Earth you can make a variety of visually pleasing, well-crafted maps with cartography or GIS software." This data set seems ideal as a replacement for the MWDBII.

Advantages:

Free / Public Domain data
Regularly updated
Wide variety of political and geographic features
Available at 3 different scales: 1:10m 1:50m and 1:110m
All feature nodes at same scale are matched
Data attributes such as country code, population, relative magnitude, etc

Disadvantages:

Is in Shapefile format which is space inefficient
No per node zoom level attribute
Different scale datasets do not match so cannot efficiently be used together for zooming

The 1:10m dataset seems ideal as the base map in Marble as it provides a higher level of detail than the current MWDBII. The 1:110m dataset seems ideal for use in a country selector widget in kdelibs. The 1:50m dataset is less detailed than the current MWDBII so may be less useful.

Using the data in the default shapefile format is not considered desirable however:

No shapefile format support in Marble (yet), would have to rely on an external library or write our own
Space inefficient (14Mb vs 2.6Mb for MWDBII)
No zoom level attribute or any node level attributes which would slow drawing
Vector level attributes are stored in .dbf format which adds complexity to implementing shapefile support

The ideal solution would seem be to convert the Natural Earth data into a more efficient file format and either calculate and store the zoom level attribute in the file or calculate it on file load. The full Natural Earth dataset would be converted, but would only ship the minimal dataset required with Marble (approx 4-5Mb?) with the remainder of the data later being downloaded via GHNS or as a separate package.

[edit] Vector Layer Improvements

The main changes required to Marble will be in the vector layer itself, removing the old PNT file vector drawing code and implementing the new dataset using the new GeoData library vector support.

Two main issues will need to be solved here

if a new file format is needed for efficient storage
if a zoom level attribute is needed for fast drawing, and if so where and how to implement the attribute

Some possibilities for the file format are:

Adapt the existing PNT format used for MWDBII to have a higher degree of accuracy and add an attributes table
Use the existing Marble serial/cache format which will be faster but the size efficiency may not be sufficient
Use shapefile by implementing a lightweight parser without dbf support but our own simple attribute table.
Implement full shapefile support including dbf (possibly using libshape)

The zoom level problem can be solved by either:

Calculate a zoom level for each point during the file conversion and save it in the new file format, but this will require more storage space.
Improve the vector drawing layer to calculate the zoom level on the fly, which would benefit all vector drawing but may be too slow.

The Douglas-Peucker algorithm may be able to be used here.

Some pros/cons to consider:

The 1:10m country file is 6.55MB and contains 533,202 points = 12.28

bytes/point compared to the PNT which is 745KB and contains 127,246 points = 5.85 bytes/point, which would suggest the NE data in PNT format would be half the size, so 6 MB in total. This could probably be further reduced by a light application of Douglas-Peucker.

The NE shapefiles have been carefully processed so shared borders and

overlapping features like rivers match exactly and other such niceties, applying the Douglas-Peucker algorithm might affect that.

A shapefile parser would allow users/apps to load other shapefiles.

We would have to reconvert and check the data every time there's a new NE

release which could be a lot of effort, but an automated shp2pnt script could prove useful to allow apps/users to display their own shapefiles in a simple way.

It is unknown if the D-H algorithm can be deployed in a way to mark each point with a detail level rather than just throwing the points away.

[edit] Calculating Zoom Level on the fly

Using GeoPainter and GeoDataLineString ("libgeodata"):

Apply Douglas-Peuker dynamically in GeoDataLineString class to set Detail Level
When GeoDataLineString modified (add/del point) set dirty flag
When node accessed for drawing, Dirty flag would trigger D-P to calculate detail level and unset dirty flag
Would benefit all vector formats, e.g. kml, ppx, shp

[edit] Possible new file format (Proposed by John)

Potential new Marble file format based on PNT:

1st integer (32 bit): Latitude in arcseconds highest bit indicates new polygon starts: header information has to be read from 3rd integer
2nd integer (32 bit): Longitude in arcseconds
Optional 3rd integer: feature key highest bit feature geometry (line or ring).

Applying this to the 1:10m dataset:

Each point takes 64 bits/8 bytes
The start of each polygon takes 96 bits
Roughly 533,202 x 8 bytes = 4 Mb for the country borders alone, not including internal border and coastline files
If that's too much to ship, then ship the 1:50m dataset as the default and download the 1:10m dataset once online

[edit] Possible more size efficient new file format (Proposed by Torsten)

For the Natural Earth Layer providing the Default data set at 0.5 arcminute resolution should be enough. This fileformat allows for even better packed data than the PNT format. For detailed polygons at arcminute scale on average it should use only 33% of the amount used by PNT.

[edit] Description of the file format

In the fileformat initally a file header is provided that provides the file format version and the number of polygons stored inside the file. A Polygon starts with the Polygon Header which provides the feature id and the number of so called "absolute nodes" that are about to follow. Absolute nodes always contain absolute geodetic coordinates. The Polygon Header also provides a flag that allows to specify whether the polygon is supposed to represent a line string ("0") or a linear ring ("1"). Each absolute node can be followed by relative nodes: These relative nodes are always nodes that follow in correct order inside the polygon after "their" absolute node. Each absolute node specifies the number of relative nodes which contain relative coordinates in reference to their absolute node. So an absolute node provides the absolute reference for relative nodes across a theoretical area of 2x2 squaredegree-area (which in practice frequently might rather amount to 1x1 square degrees).

So much of the compression works by just referencing lat/lon diffs to special "absolute nodes". Hence the compression will especially work well for polygons with many nodes with a high node density.

The parser has to convert these relative coordinates to absolute coordinates.

[edit] File Structure

File header

quint8: File format version
quint32: Number of polygons contained in the file.

Polygon Header

quint32: Feature id (either Natural Earth or Geonames).
quint32: Number of parent node chunks to follow
quint8: Flags: 1st bit: polygonIsClosed

Absolute node chunk

qint16: Latitude in halfarcminutes (allowed range = [-10800;+10800 ] halfarcminutes = [-90;+90 ] degrees )
qint16: Longitude in halfarcminutes (allowed range = [-21600;+21600 ] halfarcminutes = [-180;+180 ] degrees )
qint16: Number of child node chunks to follow (equals "0" if there are no child nodes)

Relative node chunk

qint8: Latitude-diff in halfarcminutes compared to the parent (allowed range = [-60;+60] arcminutes = [-1;+1] degrees)
qint8: Longitude-diff in halfarcminutes compared to the parent (allowed range = [-60;+60] arcminutes = [-1;+1] degrees)

[edit] Attribute Database

Metadata file:

convert / filter dbf into our own format
could just be csv or xml? or sqlite?

Rather than the Geonames ID, we could just use the Natural Earth object ID, then a look-up file/table that matches the NE ID to the ISO / FIPS / whatever code (NE provides this in the metadata) and Geonames ID (which we would have to provide). This would allow look-ups via whatever code or ID is available, and we wouldn't be reliant on Geonames IDs staying constant or being online.

[edit] Spatialite

One option would be to integrate Spatialite and use this as both the data storage for the vectors and as the attribute database. Spatialite is an extension to SQLite implementing a Spatial SQL database. Among the feature this provides is a compact data storage format and the ability to import Shapefiles and CSV files, as well as access all the standard GEOS tools if installed.

There is a 20Mb zip file available for Natural Earth in Spatialite format, it is unclear how much of Natural Earth is contained in this. A minimal dataset could be shipped by default with the full dataset downloaded later.

Spatial SQL queries could return just those vectors currently in the viewport, but repeated reloading and redrawing could be inefficient. However this may also solve the Level-of-Detail problem.

The major downside is the dependencies which include SQLite, PROJ and GEOS so on a platform like Windows would require a larger monolithic binary which defeats the purpose of shipping slimmed down data.

More research is required here. It may not be a suitable option for the default Atlas view, but would be a very powerful extension for Marble to provide lightweight GIS-like functionality.

[edit] Action Plan

A possible action plan is

Fix GeoPainter LinearRings which contain a pole not rendered correctly
Implement Douglas-Peucker reduction in GeoDataLineString
New PNT file format definition (with a different name, MBL?)
Metadata file format definition
New GeoData PNT2 file loading code (convert old data).
shp2pnt2 script to convert shp to new formats (using Perl::shp? there's shp2xxx scripts out there we could copy?), including matching to Geonames ID
split files into 'ship with', 'download asap', 'ghns'

Later add simple shapefile loading to GeoData, maybe with attibute layer?