In villandre/MRAinla: IS-MRA: An algorithm for large spatiotemporal datasets

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

Installing INLAMRA is best done with the install_git function in the remotes library, which is installed along devtools, e.g.

remotes::install_git(url = "https://github.com/villandre/MRAinla", lib = "/store/luc/R/x86_64-pc-linux-gnu-library/3.6")

The INLAMRA function lets users fit the INLA-MRA model described in Villandre et al. 2020. With adequate tuning and a large enough memory bank, they should be applicable to datasets of sizes under 1.5 million. The running time scales linearly with the required number of predictions. The software can take advantage of OpenMP parallelisation to perform that step more quickly. Enabling it is almost essential if the number of predictions to process exceeds $10,000$, cf. numOpenMPthreads in INLAMRA.control.

The arguments of INLAMRA include the usual data inputs: a numeric vector of responses, a data.frame of numeric covariates, a matrix of spatial coordinates, and a vector of time values in POSIX* or numeric format. When a POSIX* object is provided, time is scaled in days. If a numeric object is provided instead, no such assumption is made: time hyperparameters, e.g. range, will be based on the untransformed values. If predictions are required, relevant spatial and temporal coordinates, as well as covariate values, must be provided before the model is fitted. You can see all default control parameters for INLAMRA by checking ?INLAMRA.control. If the software encounters missing values, a warning is produced, and the missing observations are removed. Any missingness in the prediction input will result in NA's in the output.

In the following example, we generate seven nested grids, create $20$ knots in each subregion at depths $0$ to $6$, and keep only half the observation locations to obtain knot positions at depth $7$. We sample only $20$ points in the importance sampling step. The goal here was mainly to keep computation time reasonably low. Note that MODISdataTraining and MODISdataTest are in the STIDF format, defined in the spacetime library.

library(MRAinla)
data("MODISdataTraining")
data("MODISdataTest")
fittedModel <- INLAMRA(
  responseVec = MODISdataTraining@data$y,
  covariateFrame = subset(MODISdataTraining@data, select = -y),
  spatialCoordMat = MODISdataTraining@sp@coords,
  timePOSIXorNumericVec = time(MODISdataTraining),
  predCovariateFrame = MODISdataTest@data,
  predSpatialCoordMat = MODISdataTest@sp@coords,
  predTimePOSIXorNumericVec = time(MODISdataTest),
  control = list(
    Mlon = 3,
    Mlat = 3,
    Mtime = 1,
    numKnotsRes0 = 20,
    numValuesForIS = 10,
    tipKnotsThinningRate = 0.5
  )
)
fittedModel$hyperMarginalMoments

We can plot the predictions with the plot function, adapted to the INLAMRA class:

plot(fittedModel,
     type = "joint",
     control =  list(
       controlForScaleBar = list(
         pos = "topleft",
         cex = 2,
         pt.cex = 1.5
       ),
       controlForRasterLegend = list(
         pos = "bottomleft",
         title = "LST (Celsius)",
         bg = "white",
         cex = 2
       ),
       controlForRasterColourScale = list(
         col = "Spectral",
         breaks = 10,
         rev = TRUE,
         style = "equal",
         dec = 1,
         opacity = 0.5
       ),
       controlForRasterPlot = list(
         cex.axis = 2,
         cex.main = 3
       ),
       resolutionInMeters = 1000
     ),
)

Note that we can save the graph directly to a file by specifying a file name in argument filename. By default, the graphics engine used to export the graph is jpeg. It can be changed with the graphicsEngine option in the control parameters. We can plot marginal posteriors by specifying type = "marginals" instead.

plot(fittedModel, type = "marginals")

Notes on design

The INLA-MRA interface was designed with a focus on ease-of-use. The main function for fitting models, INLAMRA requires strictly common R objects, e.g. matrix, data.frame, list. The down side is that users can provide incompatible objects, which might break the function. INLAMRA checks for a number of common flaws in the input, but is unlikely to be able to catch everything.

Users should be aware that exceeding the amount of memory available on a given system is possible. What happens in this case is OS-specific. The software might produce a segmentation fault, which is why we recommend running it always in batch mode.

villandre/MRAinla documentation built on April 20, 2021, 7:49 a.m.