Permalink
Fetching contributors…
Cannot retrieve contributors at this time
81 lines (51 sloc) 9.8 KB

Simple Features for R

Applicant: Edzer Pebesma, Institute for Geoinformatics, University of Muenster, Germany; edzer.pebesma@uni-muenster.de

Supporting authors: Edzer Pebesma, Roger Bivand, Michael Sumner, Robert Hijmans, Virgilio Gómez-Rubio

Simple features is an open (OGC and ISO) interface standard for access and manipulation of spatial vector data (points, lines, polygons). It includes a standard SQL schema that supports storage, retrieval, query and update of feature collections via a SQL interface. All commonly used databases provide this interface. GeoJSON is a standard for encoding simple features in JSON, and is used in JavaScript and MongoDB. Well-known-text (WKT) is a text representation of simple features used often in linked data; well-known-binary ([WKB] (https://en.wikipedia.org/wiki/Well-known_text)) a standard binary representation used in databases. Simple Feature Access defines coordinate reference systems, and makes it easy to move data from longitude-latitude to projections back and forth in a standardized way.

GDAL is an open source C++ library for reading and writing both raster and vector data with more than 225 drivers (supported file formats, data base connectors, web service interfaces). GDAL is used by practically all open source geospatial projects and by many industry products (including ESRI's ArcGIS, ERDAS, and FME). It provides coordinate transformations (built on top of PROJ.4) and geometric operations (e.g. polygon intersections, unions, buffers and distance). Standards for coordinate transformations change over time; such changes are typically adopted directly in GDAL/PROJ.4 but do not easily find their way into R-only packages such as mapproj.

Since 2005, CRAN has package sp which provides classes and methods for spatial (point, line, polygon and raster) data. The approach sp takes is similar to how xts and zoo handle the time index of time series data: objects store spatial geometries separately from associated attribute data, matching by order. Package spacetime, on CRAN since 2010, extends both sp and xts to handle data that varies over both space and time.

Today, 221 CRAN packages depend on, import or link to sp, 259 when including Suggests; when including recursive dependencies these numbers are 376 and 5040. The implementation of sp does not follow simple features, but rather the practice used at the time of release, following how ESRI shapefiles are implemented. The cluster of packages around sp is shown in Andrie de Vries' blog on CRAN's network structure in green.

Off-CRAN package rgdal2 is an interface to GDAL 2.0, which uses raw pointers to interface features, but does not import any data in R, using GDAL to handle everything. CRAN Package wkb, contributed by Tibco Software, converts between WKB representations of several simple feature classes and corresponding classes in sp, and seems to be needed for Tibco software purposes.

The problem

The problems we will solve are:

  1. R can currently not represent simple features directly. It can read most simple feature classes in sp classes, but uses its own representation for this, and can only write data back without loss of information if it is furnished with ancilliary metadata encoded in a comment attribute to each Polygons object. It does for instance internally not distinguish between POLYGON and MULTIPOLYGON nor deal with several simple feature classes, including TIN and GEOMETRYCOLLECTION, nor handle CURVE geometries.
  2. The current implementation of lines and vector data in package sp is partly ambiguous (both slot ringDir or slot hole indicate whether a Polygon is a hole but are superceded by the comment attribute), complicated (to which exterior polygon does a hole belong - handled by the comment attribute), and by some considered difficult to work with (S4). The current implementation is hard to maintain because it contains incremental changes from a baseline that predated the industry-standard OGC/ISO (Simple Feature Interface Specification).
  3. The lack of support for simple features makes current interfaces to open source libraries (GDAL/OGR and PROJ.4: rgdal, GEOS: rgeos) difficult to understand and maintain, even though they work to specification.
  4. The current implementation has no scale model for coordinates.
  5. It is desirable that other R packages are offered the opportunity to migrate to more up-to-date libraries for coordinate transformations (providing proper support for datum transformation), and to avoid having to make simplifying assumptions (e.g., all spatial data come as longitude/latitude using datum WGS84; all web maps use web Mercator).

Which users will benefit from solving these problems? It will mainly affect those who use data bases or modern javascript-based web APIs which largely converged on adopting simple features (such as CartoDB), as well as those who need a simpler and more light-weight handling of spatial data in R. It will also reduce the effort for users and developers to understand the way spatial information is represented in R, making it easier to build upon and reuse the R code for this, and lead to a good, sustainable shared R code base.

In the longer run it will affect users of all packages currently reusing sp classes, when we manage to migrate sp to exclusively use the simple feature classes for representing vector data. Since the recent 2.0 release of GDAL integrates raster and vector data, having an R package that mirrors its classes makes it possible to implement operations in-database (similar to what DBI, RPostgreSQL and dplyr do), making it possible for R to manipulate spatial data that do not fit in memory.

Big Data analysis with R often proceeds by connecting R to a database that holds the data. All commonly used commercial and open source databases store spatial point, line and polygon data in the form of simple features. Representing simple features in R will simplify big data analysis for spatial data.

The plan

We want to solve the problem by carrying out the following steps (M1 refers to month 1):

  1. develop an R package that implements simple features in R, that is simple yet gives users access to the complete data, and includes an S3 representation that extends data.frame (M1-3)
  2. add to this package a C++ interface to GDAL 2.0, to read and write simple feature data, and to interface other functionality (coordinate transformation, geometry operations) (M3-8)
  3. develop and prototypically implement a migration path for sp to become compliant with simple features (M7-12)
  4. write user-oriented tutorial vignettes showing how to use it with files, data base connections, web API's, leaflet, ggmap, dplyr and so on (M7-10)
  5. write a tutorial vignette for R package writers reusing the package (M10)
  6. Collect and process community feed back (M6-12).

Failure modes and recovery plan:

  1. Failure mode: S3 classes are too simple to represent simple features class hierarchy. Recovery plan: try (i) using a list column with geometry, and nested lists to represent nested structures; (ii) use a WKT character column; (iii) using a WKB blob column

  2. Migrating sp breaks downstream packages. Recovery plan: involve Roger Bivand, Barry Rowlingson, Robert Hijmans (raster) and Tim Keitt (rgdal/rgdal2) how to proceed; be patient and smooth out problems together with package maintainers.

How can the ISC help

The following table contains the cost items.

Item Cost
employ a student assistant for one year (10 hrs/week) € 6500
one week visit of Roger Bivand to the Inst. for Geoinformatics € 1000
present the results at UseR! 2016 € 1500
Total: € 9000 (9750 USD)

The visit of Roger is anticipated halfway the project; further communications will use skype. The project has a planned duration of 12 months.

Dissemination

Development will take place on github, information will be shared and reactions and contributions invited through r-sig-geo, as well as StackOverflow and GIS StackExchange. The project will use an Apache 2.0 license for maximum dissemination (similar to GDAL, which uses X/MIT). The work will be published in 4 blogs (quarterly), announced on r-sig-geo (3300 subscribers), and intermediary results will be presented at UseR! 2016. The final result will be published in a paper either submitted to The R Journal or to the Journal of Statistical Software; this paper will be available before publication as a package vignette.

UseR! slides

UseR! 2016 slides are found here.