2014-11-06 17:06 GMT+01:00 Adam Wulkiewicz :
Hi,

I'm planning to begin the implementation of the support for geographical CS and would like to hear your opinion.

I just wanted to say that I'm glad you're working on this. I used the geographical projections from extensions in a project I finished a year ego. The project used a Boost release 1.54.0 (I think) with a snapshot of Boost.Geometry from svn trunk up front in include dirs list. To unit-test our use of the projections, we generated some test data in Matlab, and the tests confirmed accurate results of projections.

Esspecially from people already using the geographical extensions.
Basically, to support this CS we need a distance strategy and side strategy. They're already implemented (more or less) though in the extensions so not officially released yet. To release them we need to think about the interface first. The geographical strategies should have an interface similar to the one used in spherical strategies. Or the other way around, spherical strategies interface could be extended somehow to match the geographical version. In all cases the strategies should be backward compatible. Currently we have:

CARTESIAN:

- cartesian side has no ctor

side_by_triangle<> s;

- cartesian distance has no ctor

pythagoras<> s;

SPHERICAL:

- spherical side has no ctor since it doesn't need to know the radius

spherical_side_formula<> s;

- spherical distance (haversine) requires a radius of type passed to the strategy as template parameter, and by default radius = 1.0 is used (unit sphere)

haversine<RadiusType> s; // radius = 1.0 by default

so to calculate the distance on a sphere the user can pass a strategy with radius or use the default one and multiply the result by the radius:

distance(p1, p2) * radius;

GEOGRAPHICAL:

- a spherical side strategy is just used by default which I guess should be changed. I think the most precise would be a strategy finding the side using the geographic courses found using the inverse Vincenty's formula but this requires some testing
- for distance there are 2 strategies, both taking detail::ellipsoid<RadiusType>, andoyer also has a ctor taking a flattening of type RadiusType (like detail::ellipsoid<>). If 1 parameter is passed (flattening) then the radius is set to 1.0 but if 2 parameters are passed then A and B radiuses are set. Furthermore by default Earth's radiuses are used. So the situation looks like this:

vincenty<RadiusType> s(detail::ellipsoid<RadiusType>()); // A = 6378137.0, B = 6356752.314245
vincenty<RadiusType> s(detail::ellipsoid<RadiusType>(f)); // A = 1.0, B = [some fraction for oblate spheroid]
vincenty<RadiusType> s(detail::ellipsoid<RadiusType>(a, b)); // A = a, B = b
andoyer<RadiusType> s; // A = 6378137.0, B = 6356752.314245
andoyer<RadiusType> s(f); // A = 1.0, B = [some fraction for oblate spheroid]

It's confusing (this 1 and 2 paramers difference) and inconsistent with spherical strategies.
Furthermore flattening should be stored using some Floating Point type and currently it's just RadiusType. E.g. for integral RadiusType the default ctor would initialize flattening with 0.

PROPOSAL:

1. I propose to keep it simple for now, don't release detail::ellipsoid and change the default parameters to make the strategies close to the spherical ones:

vincenty<RadiusType> s(); // A = 1.0, B = [fraction], F= [fraction,default] - unit WGS84 spheroid
vincenty<RadiusType> s(a); // A = a, B = [calculated from a using default flattening], F= [fraction,default]
vincenty<RadiusType> s(a, b); // A = a, B = b, F= [calculated from a and b]

This way the calls:

distance(p1, p2) * a;

would be equivalent (assuming that RadiusType is not integral type?).

If I recall correctly, we didn't use the distance algorithm in the end. But at some prototyping stage we used distance(p1, p2, vincenty<RadiusType>()), expecting the distance on Earth. You're proposed change would break this expectation, but I don't think this is a problem.

By default FlatteningType would be:
- for floating point RadiusType <=> RadiusType
- for integral RadiusType <=> promoted to FP of the same size (int-> float, long long-> double)
- for user defined types <=> UDT
The FlatteningType could be optionally passed as a second template parameter of a strategy however I'm not sure about it since it wouldn't play nice with possible future improvement 3 (see below).

2. To decrease the probability of a mistake the strategy could take wrapped parameters (only?). This way it'd also be possible to create an ellipsoid from A,B A,F or B,F pair:

// below are functions returning wrapped values

3. Desing concepts for models parameters. And then always pass a type adapted to model's parameters concept to the strategy, instead of RadiusType. Actually a single number could be adapted to e.g. sphere parameters concept (radius) but why reserve it specifically for sphere?