Skip to content

Latest commit

 

History

History
511 lines (420 loc) · 19 KB

README.md

File metadata and controls

511 lines (420 loc) · 19 KB

gdalcubes

R-CMD-check CRAN Downloads

The R package gdalcubes aims at making analyses of large satellite image collections easier, faster, more intuitive, and more interactive.

The package represents the data as regular raster data cubes with dimensions bands, time, y, and x and hides complexities in the data due to different spatial resolutions,map projections, data formats, and irregular temporal sampling.

Features

  • Read and process multitemporal, multispectral Earth observation image collections as regular raster data cubes by applying on-the-fly reprojection, rescaling, cropping, and resampling.
  • Work with existing Earth observation imagery on local disks or cloud storage without the need to maintain a 2nd copy of the data.
  • Apply user-defined R functions on data cubes.
  • Execute data cube operation chains using parallel processing and lazy evaluation.

Among others, the package has been successfully used to process data from the Sentinel-2, Sentinel-5P, Landsat, PlanetScope, MODIS, and Global Precipitation Measurement Earth observation satellites / missions.

Installation

Install from CRAN with:

install.packages("gdalcubes")

From sources

Installation from sources is easiest with

remotes::install_git("https://github.com/appelmar/gdalcubes")

Please make sure that the git command line client is available on your system. Otherwise, the above command might not clone the gdalcubes C++ library as a submodule under src/gdalcubes.

The package builds on the external libraries GDAL, NetCDF, SQLite, and curl.

Windows

On Windows, you will need Rtools to build the package from sources.

Linux

Please install the system libraries e.g. with the package manager of your Linux distribution. Also make sure that you are using a recent version of GDAL (>2.3.0). On Ubuntu, the following commands will install all neccessary libraries.

sudo add-apt-repository ppa:ubuntugis/ppa && sudo apt-get update
sudo apt-get install libgdal-dev libnetcdf-dev libcurl4-openssl-dev libsqlite3-dev libudunits2-dev

MacOS

Using Homebrew, required system libraries can be installed with

brew install pkg-config
brew install gdal
brew install netcdf
brew install libgit2
brew install udunits
brew install curl
brew install sqlite
brew install libtiff
brew install hdf5
brew install protobuf

Getting started

Download example data

if (!dir.exists("L8_Amazon")) {
  download.file("https://hs-bochum.sciebo.de/s/8XcKAmPfPGp2CYh/download", destfile = "L8_Amazon.zip",mode = "wb")
  unzip("L8_Amazon.zip", exdir = "L8_Amazon")
}

Creating an image collection

At first, we must scan all available images once, and extract some metadata such as their spatial extent and acquisition time. The resulting image collection is stored on disk, and typically consumes a few kilobytes per image. Due to the diverse structure of satellite image products, the rules how to derive the required metadata are formalized as collection_formats. The package comes with predefined formats for some Sentinel, Landsat, and MODIS products (see collection_formats() to print a list of available formats).

library(gdalcubes)

gdalcubes_options(parallel=8)

files = list.files("L8_Amazon", recursive = TRUE, 
                   full.names = TRUE, pattern = ".tif") 
length(files)
## [1] 1805
sum(file.size(files)) / 1024^2 # MiB
## [1] 1919.12
L8.col = create_image_collection(files, format = "L8_SR", out_file = "L8.db")
L8.col
## Image collection object, referencing 180 images with 10 bands
## Images:
##                                       name      left       top    bottom
## 1 LC08_L1TP_226063_20140719_20170421_01_T1 -54.15776 -3.289862 -5.392073
## 2 LC08_L1TP_226063_20140820_20170420_01_T1 -54.16858 -3.289828 -5.392054
## 3 LC08_L1GT_226063_20160114_20170405_01_T2 -54.16317 -3.289845 -5.392064
## 4 LC08_L1TP_226063_20160724_20170322_01_T1 -54.16317 -3.289845 -5.392064
## 5 LC08_L1TP_226063_20170609_20170616_01_T1 -54.17399 -3.289810 -5.392044
## 6 LC08_L1TP_226063_20170711_20170726_01_T1 -54.15506 -3.289870 -5.392083
##       right            datetime        srs
## 1 -52.10338 2014-07-19T00:00:00 EPSG:32622
## 2 -52.11418 2014-08-20T00:00:00 EPSG:32622
## 3 -52.10878 2016-01-14T00:00:00 EPSG:32622
## 4 -52.10878 2016-07-24T00:00:00 EPSG:32622
## 5 -52.11958 2017-06-09T00:00:00 EPSG:32622
## 6 -52.09798 2017-07-11T00:00:00 EPSG:32622
## [ omitted 174 images ] 
## 
## Bands:
##         name offset scale unit       nodata image_count
## 1    AEROSOL      0     1                           180
## 2        B01      0     1      -9999.000000         180
## 3        B02      0     1      -9999.000000         180
## 4        B03      0     1      -9999.000000         180
## 5        B04      0     1      -9999.000000         180
## 6        B05      0     1      -9999.000000         180
## 7        B06      0     1      -9999.000000         180
## 8        B07      0     1      -9999.000000         180
## 9   PIXEL_QA      0     1                           180
## 10 RADSAT_QA      0     1                           180

Creating data cubes

To create a regular raster data cube from the image collection, we define the geometry of our target cube as a data cube view, using the cube_view() function. We define a simple overview, covering the full spatiotemporal extent of the imagery at 1km x 1km pixel size where one data cube cell represents a duration of one year. The provided resampling and aggregation methods are used to spatially reproject, crop, and rescale individual images and combine pixel values from many images within one year respectively. The raster_cube() function returns a proxy object, i.e., it returns immediately without doing any expensive computations.

v.overview = cube_view(extent=L8.col, dt="P1Y", dx=1000, dy=1000, srs="EPSG:3857", 
                      aggregation = "median", resampling = "bilinear")
raster_cube(L8.col, v.overview)
## A data cube proxy object
## 
## Dimensions:
##                 low              high count pixel_size chunk_size
## t        2013-01-01        2019-12-31     7        P1Y          1
## y -764014.387686915 -205014.387686915   559       1000        192
## x -6582280.06164712 -5799280.06164712   783       1000        192
## 
## Bands:
##         name offset scale nodata unit
## 1    AEROSOL      0     1    NaN     
## 2        B01      0     1    NaN     
## 3        B02      0     1    NaN     
## 4        B03      0     1    NaN     
## 5        B04      0     1    NaN     
## 6        B05      0     1    NaN     
## 7        B06      0     1    NaN     
## 8        B07      0     1    NaN     
## 9   PIXEL_QA      0     1    NaN     
## 10 RADSAT_QA      0     1    NaN

Processing data cubes

We can apply (and chain) operations on data cubes:

x = raster_cube(L8.col, v.overview) |>
  select_bands(c("B02","B03","B04")) |>
  reduce_time(c("median(B02)","median(B03)","median(B04)"))
x
## A data cube proxy object
## 
## Dimensions:
##                 low              high count pixel_size chunk_size
## t        2013-01-01        2019-12-31     1        P7Y          1
## y -764014.387686915 -205014.387686915   559       1000        192
## x -6582280.06164712 -5799280.06164712   783       1000        192
## 
## Bands:
##         name offset scale nodata unit
## 1 B02_median      0     1    NaN     
## 2 B03_median      0     1    NaN     
## 3 B04_median      0     1    NaN
plot(x, rgb=3:1, zlim=c(0,1200))

library(RColorBrewer)
 raster_cube(L8.col, v.overview) |>
  select_bands(c("B04","B05")) |>
  apply_pixel(c("(B05-B04)/(B05+B04)"), names="NDVI") |>
  plot(zlim=c(0,1),  nbreaks=10, col=brewer.pal(9, "YlGn"), key.pos=1)

Calling data cube operations always returns proxy objects, computations are started lazily when users call e.g. plot().

Animations

Multitemporal data cubes can be animated (thanks to the gifski package):

v.subarea.yearly = cube_view(extent=list(left=-6180000, right=-6080000, bottom=-550000, top=-450000, 
                             t0="2014-01-01", t1="2018-12-31"), dt="P1Y", dx=50, dy=50,
                             srs="EPSG:3857", aggregation = "median", resampling = "bilinear")

raster_cube(L8.col, v.subarea.yearly) |>
  select_bands(c("B02","B03","B04")) |>
  animate(rgb=3:1,fps = 2, zlim=c(100,1000), width = 400, 
          height = 400, save_as = "man/figures/animation.gif")

Data cube export

Data cubes can be exported as single netCDF files with write_ncdf(), or as a collection of (possibly cloud-optimized) GeoTIFF files with write_tif(), where each time slice of the cube yields one GeoTIFF file. Data cubes can also be converted to terra or starsobjects:

raster_cube(L8.col, v.overview) |>
  select_bands(c("B04","B05")) |>
  apply_pixel(c("(B05-B04)/(B05+B04)"), names="NDVI") |>
  write_tif() |>
  terra::rast() -> x
x
## class       : SpatRaster 
## dimensions  : 559, 783, 7  (nrow, ncol, nlyr)
## resolution  : 1000, 1000  (x, y)
## extent      : -6582280, -5799280, -764014.4, -205014.4  (xmin, xmax, ymin, ymax)
## coord. ref. : WGS 84 / Pseudo-Mercator (EPSG:3857) 
## sources     : cube_845e62ea0e0a2013-01-01.tif  
##               cube_845e62ea0e0a2014-01-01.tif  
##               cube_845e62ea0e0a2015-01-01.tif  
##               ... and 4 more source(s)
## names       : NDVI, NDVI, NDVI, NDVI, NDVI, NDVI, ...
raster_cube(L8.col, v.overview) |>
  select_bands(c("B04","B05")) |>
  apply_pixel(c("(B05-B04)/(B05+B04)"), names="NDVI") |>
  stars::st_as_stars() -> y
y
## stars object with 3 dimensions and 1 attribute
## attribute(s), summary of first 1e+05 cells:
##             Min.   1st Qu.   Median      Mean  3rd Qu.      Max.  NA's
## NDVI  -0.5595199 0.4207425 0.723503 0.5765454 0.849606 0.8892204 79500
## dimension(s):
##      from  to   offset delta                   refsys point
## x       1 783 -6582280  1000 WGS 84 / Pseudo-Mercator    NA
## y       1 559  -205014 -1000 WGS 84 / Pseudo-Mercator    NA
## time    1   7       NA    NA                  POSIXct FALSE
##                                                   values x/y
## x                                                   NULL [x]
## y                                                   NULL [y]
## time [2013-01-01,2014-01-01),...,[2019-01-01,2020-01-01)

To reduce the size of exported data cubes, compression and packing (conversion of doubles to smaller integer types) are supported.

If only specific time slices of a data cube are needed, select_time() can be called before plotting / exporting.

raster_cube(L8.col, v.overview) |>
  select_bands(c("B04","B05")) |>
  apply_pixel(c("(B05-B04)/(B05+B04)"), names="NDVI") |>
  select_time(c("2015", "2018")) |>
  plot(zlim=c(0,1), nbreaks=10, col=brewer.pal(9, "YlGn"), key.pos=1)

User-defined functions

Users can pass custom R functions to reduce_time() and apply_pixel(). Below, we derive a greenest pixel composite by returning RGB values from pixels with maximum NDVI for all pixel time-series.

v.subarea.monthly = cube_view(view = v.subarea.yearly, dt="P1M", dx = 100, dy = 100,
                              extent = list(t0="2015-01", t1="2018-12"))
raster_cube(L8.col, v.subarea.monthly) |>
  select_bands(c("B02","B03","B04","B05")) |>
  apply_pixel(c("(B05-B04)/(B05+B04)"), names="NDVI", keep_bands=TRUE) |>
  reduce_time(names=c("B02","B03","B04"), FUN=function(x) {
    if (all(is.na(x["NDVI",]))) return(rep(NA,3))
    return (x[c("B02","B03","B04"), which.max(x["NDVI",])])
  }) |>
  plot(rgb=3:1, zlim=c(100,1000))

Extraction of pixels, time series, and summary statistics over polygons

In many cases, one is interested in extracting sets of points, time series, or summary statistics over polygons, e.g., to generate training data for machine learning models. Package version 0.6 therefore introduces the extract_geom() function, which replaces the previous implementations in query_points(), query_timeseries(), and zonal_statistics().

Below, we randomly select 100 locations and query values of single data cube cells and complete time series.

x = runif(100, v.overview$space$left, v.overview$space$right)
y = runif(100, v.overview$space$bottom, v.overview$space$top)
t = sample(as.character(2013:2019), 100, replace = TRUE)
df = sf::st_as_sf(data.frame(x = x, y = y), coords = c("x", "y"), crs = v.overview$space$srs)

# spatiotemporal points
raster_cube(L8.col, v.overview) |>
  select_bands(c("B04","B05")) |>
  extract_geom(df, datetime = t) |>
  dplyr::sample_n(15) # print 15 random rows
##    FID       time       B04       B05
## 21  95 2016-01-01  182.3935 3360.3492
## 50  11 2016-01-01  282.0869 3039.4177
## 34  96 2018-01-01  885.3366 3565.0468
## 54   4 2019-01-01  171.4910 2825.5037
## 38   1 2018-01-01  249.7769 3091.6986
## 42  64 2018-01-01  315.9540 3326.0070
## 3   18 2014-01-01  720.9067 3689.0444
## 55  47 2019-01-01  569.0251 2844.1652
## 22  74 2017-01-01  264.0236 3036.4862
## 29  73 2017-01-01  198.0629 3135.8718
## 39  38 2018-01-01  201.2096 2882.1543
## 28  30 2017-01-01  171.2704 2754.2129
## 61  27 2019-01-01  405.6078  588.0934
## 51  25 2019-01-01  150.7253 2886.3868
## 25  19 2016-01-01 3593.0970 5285.5944
# time series at spatial points
raster_cube(L8.col, v.overview) |>
  select_bands(c("B04","B05")) |>
  extract_geom(df) |>
  dplyr::sample_n(15) # print 15 random rows
##     FID       time      B04      B05
## 319  98 2017-01-01 217.7226 3296.470
## 25   86 2013-01-01 199.9388 2844.481
## 100  58 2014-01-01 202.8860 2869.232
## 256  43 2017-01-01 280.8320 3187.573
## 390  41 2019-01-01 149.7427 2879.540
## 60   45 2013-01-01 239.1001 3219.560
## 288  85 2016-01-01 309.2750 2876.053
## 290  24 2017-01-01 238.6707 3151.653
## 135  31 2015-01-01 951.8869 3004.181
## 315   4 2017-01-01 146.7365 2891.950
## 66   18 2015-01-01 436.6083 3535.842
## 381  96 2019-01-01 190.6946 2812.518
## 40   49 2013-01-01 169.4907 2761.769
## 284  33 2016-01-01 225.0206 2925.426
## 222  45 2016-01-01 295.2418 3153.687

In the following, we use the example Landsat dataset (reduced resolution) from the package and compute median NDVI values within some administrative regions in New York City. The result is a data.frame containing data cube bands, feature IDs, and time as columns.

L8_files <- list.files(system.file("L8NY18", package = "gdalcubes"),
                       ".TIF", recursive = TRUE, full.names = TRUE)
v = cube_view(srs="EPSG:32618", dy=300, dx=300, dt="P1M", 
              aggregation = "median", resampling = "bilinear",
              extent=list(left=388941.2, right=766552.4,
                          bottom=4345299, top=4744931, 
                          t0="2018-01", t1="2018-12"))
sf = sf::st_read(system.file("nycd.gpkg", package = "gdalcubes"), quiet = TRUE)

raster_cube(create_image_collection(L8_files, "L8_L1TP"), v) |>
  select_bands(c("B04", "B05")) |>
  apply_pixel("(B05-B04)/(B05+B04)", "NDVI") |>
  extract_geom(sf, FUN = median) -> zstats

dplyr::sample_n(zstats, 15) # print 15 random rows
##    FID       time         NDVI
## 1   68 2018-03-01  0.008483257
## 2   49 2018-05-01  0.041807600
## 3   41 2018-01-01 -0.012513485
## 4   47 2018-10-01  0.002116781
## 5   71 2018-08-01  0.255297575
## 6   26 2018-12-01  0.060638615
## 7   48 2018-06-01  0.055183957
## 8   56 2018-10-01  0.138487053
## 9   25 2018-10-01  0.093297342
## 10  58 2018-04-01  0.022237839
## 11  57 2018-12-01  0.053809129
## 12  43 2018-08-01  0.065916489
## 13   2 2018-08-01  0.130797459
## 14  31 2018-10-01  0.044141370
## 15  32 2018-01-01  0.058137169

We can combine the result with the original features by a table join on the FID column using merge():

sf$FID = rownames(sf)
x = merge(sf, zstats, by = "FID")
plot(x[x$time == "2018-07-01", "NDVI"])

When using input features with additional attributes / labels, the extract_geom() function hence makes it easy to create training data for machine learning models.

More Features

Cloud support with STAC: gdalcubes can be used directly on cloud computing platforms including Amazon Web Services, Google Cloud Platform, and Microsoft Azure. Imagery can be read from their open data catalogs and discovered by connecting to STAC API endpoints using the rstac package (see links at the end of this page).

Machine learning: The built-in functions extract_geom and predict help to create training data and apply predictions on data cubes using machine learning models as created from packages caret or tidymodels.

Further operations: The previous examples covered only a limited set of built-in functions. Further data cube operations for example include spatial and/or temporal slicing (slice_time, slice_space), cropping (crop), application of moving window / focal operations (window_time, window_space), filtering by arithmetic expressions on pixel values and spatial geometries (filter_pixel, filter_geom), and combining two or more data cubes with identical shape (join_bands).

Further reading