Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

https://github.com/appelmar/gdalcubes

Creating and analyzing Earth observation data cubes in R
https://github.com/appelmar/gdalcubes

remote-sensing rstats satellite-imagery spatial-analysis

Last synced: about 1 month ago
JSON representation

Creating and analyzing Earth observation data cubes in R

Lists

README 

        
---

output: github_document

#always_allow_html: true

---



```{r, include=FALSE}

knitr::opts_chunk$set(echo = TRUE)

knitr::opts_chunk$set(fig.path = "man/figures/")

```



# gdalcubes 



[![R-CMD-check](https://github.com/appelmar/gdalcubes_R/actions/workflows/R-CMD-check.yml/badge.svg)](https://github.com/appelmar/gdalcubes_R/actions/workflows/R-CMD-check.yml)

[![CRAN](https://www.r-pkg.org/badges/version/gdalcubes)](https://cran.r-project.org/package=gdalcubes)

[![Downloads](https://cranlogs.r-pkg.org/badges/grand-total/gdalcubes)](https://cran.r-project.org/package=gdalcubes)



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, Landsat, PlanetScope, MODIS, and Global Precipitation Measurement Earth observation satellites / missions.



# Installation



Install from CRAN with:



```{r, eval=FALSE}

install.packages("gdalcubes")

```



## From sources



Installation from sources is easiest with



```{r, eval=FALSE}

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

```



Please make sure that the [git command line client](https://git-scm.com/downloads) 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](https://www.gdal.org), [NetCDF](https://www.unidata.ucar.edu/software/netcdf), [SQLite](https://www.sqlite.org), and [curl](https://curl.haxx.se/libcurl). 



## Windows



On Windows, you will need [Rtools](https://cran.r-project.org/bin/windows/Rtools). System libraries are automatically downloaded from [rwinlib](https://github.com/rwinlib).



## 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 install all 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

Use [Homebrew](https://brew.sh) to install system libraries with



```

brew install pkg-config

brew install gdal

brew install netcdf

brew install libgit2

brew install udunits

brew install curl

brew install sqlite

```



# Getting started



## Download example data

```{r download}

if (!dir.exists("L8_Amazon")) {

  download.file("http://data.gdalcubes.org/L8_Amazon.zip", 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).  



```{r}

library(gdalcubes)



gdalcubes_options(parallel=8)



files = list.files("L8_Amazon", recursive = TRUE, 

                   full.names = TRUE, pattern = ".tif") 

length(files)

sum(file.size(files)) / 1024^2 # MiB



L8.col = create_image_collection(files, format = "L8_SR", out_file = "L8.db")

L8.col

```



## 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.



```{r}

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)

```



## Processing data cubes



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



```{r cubes}

x = raster_cube(L8.col, v.overview) |>

  select_bands(c("B02","B03","B04")) |>

  reduce_time(c("median(B02)","median(B03)","median(B04)"))

x



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](https://cran.r-project.org/package=gifski)):



```{r animation,results='hide'}

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")

```



![](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 `raster` or `stars`objects:



```{r extpkgload, include=FALSE} 

library(raster)

library(stars)

```



```{r, warning=FALSE}

raster_cube(L8.col, v.overview) |>

  select_bands(c("B04","B05")) |>

  apply_pixel(c("(B05-B04)/(B05+B04)"), names="NDVI") |>

  write_tif() |>

  raster::stack() -> x

x



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

```



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.



```{r select_time}

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.



```{r greenest_pixel_composite}

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.



```{r extract}

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



# 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

```



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.



```{r zonal_statistics}

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

```



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



```{r}

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](https://cran.r-project.org/package=rstac) (see links at the end of this page).



**Masks**: Mask bands (e.g. general pixel quality measures or cloud masks) can be applied during the construction of the raster data cube, such that masked values will not contribute to the data cube values. 



**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`), apply moving window filters over time series (`window_time`), 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`).



# Limitations



* Data cubes are limited to four dimensions ([stars](https://cran.r-project.org/package=stars) has cubes with any number of dimensions). 

* Some operations such as `window_time()` do not support user-defined R functions at the moment.

* Images must be orthorectified / regularly gridded; there is no support for curvilinear grids.

* There is no support for vector data cubes ([stars](https://cran.r-project.org/package=stars) has vector data cubes). 



# Further reading



* [Official R package website](https://gdalcubes.github.io)

* [Tutorial on YouTube](https://youtu.be/Xlg__2PeTXM?t=3693) how to use gdalcubes in the cloud, streamed at OpenGeoHub Summer School 2021

* [1st blog post on r-spatial.org](https://www.r-spatial.org/r/2019/07/18/gdalcubes1.html) 

* [2nd blog post on r-spatial.org](https://r-spatial.org/r/2021/04/23/cloud-based-cubes.html) describing how to use gdalcubes in cloud-computing environments

* [Open access paper](https://www.mdpi.com/2306-5729/4/3/92) in the special issue on Earth observation data cubes of the data journal