Plotting interpolated data on map

Question

I have survey data of species richness that was taken at various sites in the Chesapeake Bay, USA, and I would like to graphically present the data as a \"heat map.\"

<

Answer 1

I have a number of remarks on your post:

Using kriging

I see that you are using geostatistics to construct your heatmap. You could also consider other interpolation techniques such as splines (e.g. Thin plate splines in the fields package). These make less assumptions about the data (e.g. stationarity), and can also visualize your data just fine. The reduction in the number of assumptions might help in case you send it to a journal, then you have less to explain to the reviewers. You can also compare a few interpolation techniques if you want, see a report I wrote for some tips.

Data projection

I see that you are using lat long coordinates for kriging. Edzer Pebesma (author of gstat) remarked that there are no variogram models that are suitable for lat lon coordinates. This is because in lat lon the distances are not straight (i.e. Euclidean), but over a sphere, (i.e. Great circle distances). There are no covariance functions (or variogram models) that are valid for spherical coordinates. I recommend projecting them using spTransform from the rgdal package before using automap.

The rgdal package uses the proj4 projection library to perform the calculations. To project your data you first need to define its projection:

proj4string(df) = "+proj=longlat +ellps=WGS84 +datum=WGS84 +no_defs"

The proj4 string on the right hand side of the expression above defines the type of projection (+proj), the ellips that was used (+ellps) and the datum (+datum). To understand what these terms mean, you have to imagine the Earth as a potato. The Earth is not perfectly spherical, this is defined by the ellips. Neither is the Earth a perfect ellipsoid, but the surface is more irregular. This irregularity is defined by the datum. See also this article on Wikipedia.

Once you have the projection defined, you can use spTransform:

project_df = spTransform(df, CRS("+proj= etcetc"))

where CRS("+proj etc") defines the target projection. Which projection is appropriate depends on your geographical location and the size of your study area.

Plotting with ggplot2

For adding polygons or polylines to ggplot, please to a look the documentation of coord_map. This includes an example of using the maps package to plot country boundaries. If you need to load for example shapefiles for your study area, you can do so using rgdal. Do remember that ggplot2 works with data.frame's, not SpatialPolygons. You can transform SpatialPolygons to data.frame using:

poly_df = fortify(poly_Spatial)

See also this function I created to plot spatial grids. It works directly on SpatialGrids/Pixels. Note that you need to source one or two additional files from that repository (continuousToDiscrete).

Creating interpolation grid

I created automap to generate an output grid when none was specified. This is done by creating a convex hull around the data points, and sampling 5000 points inside it. The boundaries of the prediction area, and the number of points sampled in it (and thus the resolution) is quite arbitrary. For a specific application the shape of the prediction area can be derived from a polygon, using spsample to sample points inside the polygon. How many points to sample, and thus the resolution, depends on two things:

• the kind of data you have, For example, if your data is very smooth, there is not much point in raising the resolution really high in comparison to this smoothness. Alternatively, if your data has many small scale strcutures, you need a high resolution. This is only possible ofcourse if you have the observations to support this high resolution.
• the density of data. If your data is more dense, you can raise the resolution.

If you use your interpolated map for subsequent analyses, getting the resolution right is important. If you use the map purely for visuatlisation purposes, this is less important. Note however that in both cases a too high resolution can be misleading as to the accuracy of your predictions, and that a too low resolution does not do justice to the data.