Temporal data processing/GRASS R raster time series processing

From GRASS-Wiki
< Temporal data processing
Revision as of 06:16, 10 September 2016 by Veroandreo (talk | contribs) (adding figs and outputs)

Jump to: navigation, search

GRASS-R / R-GRASS for raster time series processing

This little example will guide you through the steps to export a Spatio-Temporal Raster Dataset (strds) stored in GRASS, import it into R, prepare the data properly to use the Data INterpolation Empirical Orthogonal Functions algorithm (DINEOF) and, after running it, rebuild your raster time series, export it and import the new strds into GRASS.

We will use North Carolina climatic data already available as a GRASS location. You can get it here. If not done yet, you need to unzip the file and paste into your GRASS database folder (usually named grassdata).

Shall we start?

1. Open GRASS in the location/mapset of interest

grass70 $HOME/grassdata/nc_climate_spm_2000_2012/climate_1970_2012/
  • List the available raster maps
g.list type=raster
g.list type=raster pattern="*tempmean"
  • Create strds with mean temperature maps: t.create
t.create output=tempmean type=strds temporaltype=absolute \
    title="Average temperature" description="Monthly temperature average in NC [deg C]"
  • Register maps in the strds (t.register), list the strds in the mapset (t.list) and obtain general info for a particular space-time dataset (t.info). More details about different options to register maps in strds can be found in the map registration wiki.
t.register -i input=tempmean type=raster start=2000-01-01 increment="1 months" \
    maps=`g.list type=raster pattern="*tempmean" separator=comma` 

t.list type=strds

t.info input=tempmean
t.rast.list input=tempmean

t.rast.univar input=tempmean

As said before, We will use DINEOF as gap-filling technique. This method is not (yet) available in GRASS GIS, so we need to export the raster time series.

# set default region
g.region -d

t.rast.export input=tempmean output=tempmean4R.tar.gzip compression=gzip

2. Moving to R

From the GRASS console, open R (or rstudio)

rstudio $HOME/Documents/foss4g_bonn/ts_grass_r &

Note that you can directly open a project in which you were already working (as in the example above) or just type "R" or "rstudio" followed by "&" so you do not loose the command prompt in GRASS. More details and examples of how to use R or rstudio within a GRASS session can be found in the dedicated wiki: look for R within GRASS and Rstudio within GRASS.

Already in R, we first load "rgrass7", the library that provides the interface with GRASS GIS 7, check some basic information about the session and list maps.

> library(rgrass7)
# check GRASS environment
> genv <- gmeta()
> genv
gisdbase    /home/veroandreo/grassdata 
location    nc_climate_spm_2000_2012 
mapset      climate_1970_2012 
rows        104 
columns     139 
north       260500 
south       208500 
west        624500 
east        694000 
nsres       500 
ewres       500 
projection  +proj=lcc +lat_1=36.16666666666666 +lat_2=34.33333333333334 +lat_0=33.75
+lon_0=-79 +x_0=609601.22 +y_0=0 +no_defs +a=6378137 +rf=298.257222101
+towgs84=0.000,0.000,0.000 +to_meter=1 

This gives us information about projection and region settings, among other things, that we will need afterwards (see below).

# list maps
> execGRASS("g.list", parameters = list(type = "raster", pattern="*temp*"))

Given that neither GRASS temporal modules nor GRASS space time data sets are (yet) enabled/implemented in R, when you need to process a time series from GRASS you need to export all maps and then read them into R. Therefore, we need to load the following packages, too:

> library(spacetime)
> library(raster)
> library(rgdal)

Now, we read the strds exported from GRASS into R

> tempmean_in_R <- read.tgrass("tempmean4R.tar.gzip", localName = FALSE, useTempDir = FALSE)

Note that R imports our exported strds as a RasterStack.

Let us see if all is there:

# get information about the RasterStack
> class(tempmean_in_R)
[1] "RasterStack"
[1] "raster"
> dim(tempmean_in_R)
[1] 104 139 156

We can also ask summaries from particular layers in the stack or of the whole RasterStack

# summary of the first layer
> summary(tempmean_in_R1)
# summary of the RasterStack
> summary(tempmean_in_R)

And, of course, we may also plot a certain layer

# plot 3rd layer in the stack
> plot(tempmean_in_R,3) 

For most methods in R (For example: Theil-Sen slope, fft, dineof), we will need to transform our RasterStack into a matrix or data.frame. In the particular case of this example, we need to transform our RasterStack into a matrix mxt, with m=maps and t=time, i.e.: in this matrix we have our maps as rows and time in columns. Therefore, in each column we have the time series for a given pixel. Let us start, then.

# rasterstack to matrix (this gives us maps as columns and time in rows)
> txm_tempmean <- as.matrix(tempmean_in_R) 
> dim(txm_tempmean)
[1] 14456   156
# transpose matrix (we need time in columns and maps unfolded in rows)
> mxt_tempmean <- t(txm_tempmean)
> dim(mxt_tempmean)
[1]   156 14456

All the point of DINEOF is gap-filling. However, as our sample dataset comes from interpolations of weather stations, we do not have such gaps... So, for this example, we need to create them.

# Check for gaps in data
> sum(is.na(mxt_tempmean))
# Create some holes. 
# Total pixel counts: 14456*156=2255136
# NULL values must appear as NaN
> set.seed(46)
> n = 400000
> mxt_tempmean[mysamples <- sample(length(mxt_tempmean), n)] <- NaN
> sum(is.na(mxt_tempmean))
[1] 400000

Now we are ready, to use DINEOF to fill our gappy data. You can find more info about DINEOF in the official site. However, the R package providing DINEOF is called sinkr and can be installed from here.

# Load library containing dineof function and other required libraries
> library(sinkr)
> library(irlba)
> library(Matrix)
# or, download the dineof function from the repository and source it:
# source('dineof.r')
# Run the algorithm - default settings 
> result_tempmean <- dineof(mxt_tempmean) 

The result_tempmean object is a list with all results from dineof. You may investigate yourself, try with other settings and compare.

# Explore what's inside
> names(result_tempmean)
[1] "Xa"    "n.eof" "RMS"   "NEOF" 

For the purpose of this example, we will only extract the reconstructed spatio-temporal matrix and display it:

# Extract gap-filled matrix
> tempmean_new <- result_tempmean$Xa 
# Set color palette
> library(RColorBrewer)
> pal <- colorRampPalette(c("blue", "cyan", "yellow", "red"))
# Display the gappy and gap-filled spatio-temporal matrixes
> par(mfrow = c(2,1))
> image(mxt_tempmean, col = pal(100), main = "Original data with gaps")
> image(tempmean_new, col = pal(100), main = "Gap-filled data")
Plot gaps nogaps.png
# Plot the time series of a single pixel from both matrixes 
> pixel <- 10000
> ylim <- range(c(mxt_tempmean[,pixel], tempmean_new[,pixel]), na.rm = TRUE)
> plot(mxt_tempmean[,pixel], t = "l", ylim = ylim, lwd = 2, ylab = "Temperature", xlab = "Time")
> lines(tempmean_new[,pixel], col = 2, lty = 3)
> legend(123, 13, c("original","gap-filled"), lty = c(1,3), lwd = c(2,1), col = c(1,2), cex = 0.8)
> abline(h = 0, v = 0, col = "gray60")
Time series.png

Let us assume that we are happy with the result, we need now to go back to a strds format. We need to rebuild the raster time series starting from a matrix mxt. We first create empty rasters of the same dimensions of our original data and then, we fill them with each row of the gap-filled matrix. What we obtain is a list of rasters.

# Create raster objects to fill with data from tempmean_new
> tempmean_2_grass <- raster(nrows = 104, ncols = 139, 
           xmn = 624500, xmx = 694000, 
           ymn = 208500, ymx = 260500, 
           crs = tempmean_in_R)
> dim(tempmean_2_grass)
[1] 104 139   1
# Extract data from each row of Xa and build up a raster layer
# Output: list of rasters
> tempmean_new_rl <- lapply(1:nrow(tempmean_new), function(i) {
  setValues(tempmean_2_grass, tempmean_new[i,])
  } )
> length(tempmean_new_rl)
[1] 156
> dim(tempmean_new_rl3)
[1] 104 139   1
# Plot a raster layer in the list
> plot(tempmean_new_rl3)
# Compare with layer 3 in the original data
> plot(tempmean_in_R,3)

So far we have a list of rasters. But to export them and import them back into GRASS, we need again a RasterStack.

# rebuild RasterStack
> tempmean_new_rs <- stack(tempmean_new_rl) 
> class(tempmean_new_rs)
# we need to add time to the RasterStack
> time_4_new_rs <- seq(as.Date("2001-01-01"), as.Date("2012-12-01"))
> tempmean_new_rs_and_time <- setZ(tempmean_new_rs, time_4_new_rs, name="time")

Finally, we export RasterStack to GRASS with write.tgrass

> write.tgrass(tempmean_new_rs, "tempmean_new_from_R.tar.gzip")

3. Back to GRASS

Now, switch back to GRASS console and import the strds with t.rast.import. This command will create a strds and register all maps in it.

t.rast.import input=tempmean_new_from_R.tar.gz output=tempmean_dineof base=tempmean_dineof extrdir=/tmp