Exploring Land Cover Data (Impact Observatory)

Context¶

Purpose¶

Introduce manipulation and exploratory analysis of classified land use and cover data, using example data created by Impact Observatory from ESA Sentinel-2 imagery.

Dataset description¶

There are now many classified (categorical) land cover data products freely available that are useful for Environmental Data Science. These include:

• ESA CCI land cover, 300m spatial resolution global extent for years 1992-2015
• Copernicus Global Land Cover, 100m spatial resolution global extent for years 2015-2019
• USGS LCMAP, 30m spatial resolution for USA for years 1985-2020
• UKCEH LCMs, various spatial resolutions for UK for various years 1990-2020
• mapbiomas, 30m spatial resolution for Brazil for years 1985-2020
• Impact Observatory, 10m spatial resolution global extent for 2017-2021

These products are provided as 2D rasters (spatial) or 3D data cubes (spatio-temporal). The number and classification of discrete land cover classes varies between products, but at their most basic will distinguish between broad land covers such as ‘crops’, ‘forest’ and ‘built-up’. The nominal (categorical) character of the data influences the types of analysis appropriate.

This notebook uses data created by Impact Observatory. The data are a time series for 2017-2021 of annual global land use and land cover (LULC) mapped at 10m spatial resolution Observatory, 2022. The data are derived from ESA Sentinel-2 imagery with each annual map specifying individual pixels as belonging to one of 9 LULC classes Karra et al., 2021. The Impact Observatory LULC model uses deep learning methods to infer a single annual LULC class for each pixel in a Sentinel-2 image. Each annual global LULC map is produced by aggregating multiple inferences for images from across a given year (requiring processing approximately 2 million images to create each annual map).

Highlights¶

• Fetch land cover product data from an online repository
• Visualise the data (maps and spatio-temporal plots)
• Analyse aggregate change through bar charts and similar visualisation
• Analyse pixel-by-pixel change including use of sankey diagrams
• Analyse zonal change using ancillary vector data

Contributions¶

Dataset originator/creator¶

• Esri (licensor)
• Impact Observatory (processor, producer, licensor)
• Microsoft (host)

The data are available under a Creative Commons BY-4.0 license.

Code¶

• Data loading code built on snippet from acocac via ODC.stac docs and a Microsoft Planetary example notebook.

Load libraries¶

Set project structure¶

The next cell creates a separate folder to save any notebook outputs you should wish (e.g. below we use this directory in the demo of how you might load data from a local file).

1. Fetch and Load Data¶

This notebook uses annual land cover data from Impact Observatory, 10m spatial resolution and global extent for 2017-2021. These data are hosted online by Microsoft’s Planetary Computer which enable you, the user, to run the code in the cloud (e.g. via binder).

Below we will use the pystac-client package to access STAC information from the Planetary Computer API, then read it using the odc-stac package. In doing so we will:

• fetch data for a study area in Mato Grosso, Brazil
• resample the original 10m spatial resolution to 300m (to decrease execution times in this example notebook)

Cloud Data¶

First we need to specify the location and extent of study area to fetch the relevant data. We do this by specying a bounding box around the center point of the study area.

We can then visualise the location and extent of the study area using hvPlot.

With our study area defined, we can now query Microsoft Planetary Computer data that is available for this location using pystac-client.

For the available datasets we can access and view the associated metadata.

The global LULC datasets are for ‘tiles’ covering regions of the earth’s surface - each line in the DataFrame above is for one year for tile 21L. We can also plot to show the study area does not cover multiple land cover tiles.

From the figure above we can see the whole study area (represented by the red rectangle) is within in the tiles (represented by the blue rectangles) returned by the pystac query.

Now we are happy we know what datasets we wish to fetch, we can use odc-stac to fetch and load the data into memory, specifying the projection and spatial resolution we want as well as how we want data grouped in layers:

The data object we have created is an xarray Dataset which is a dict-like collection of DataArray objects with aligned dimensions. We can check the dimensions, coordinates, data and attributes (see xarray terminology):

Load Local Data¶

In the previous section we fetched data from an online repository and loaded it into memory. In the sections below we will continue to work with that (and you can skip to the next section now, if you don’t want to see how to save and/or read data from a local drive).

The recommended way to store xarray data structures is netCDF, a binary file format for self-described datasets that originated in the geosciences.

For example, to save the xarray Dataset we created above on disk as netCDF, we would use something like:

lcxr.to_netcdf(os.path.join(notebook_folder,"MT-study-area.nc"))

Then to load the local netCDF file into a xarray Dataset we would use:

lcxr_disk = xr.open_dataset(os.path.join(notebook_folder,"MT-study-area.nc"), decode_coords="all")

2. Visualisation¶

Our dataset contains a 2D raster layer of land cover types for individual snapshots in time. Each map is a composite of land use/cover predictions for 9 classes throughout the year in order to generate a representative snapshot of each year.

We can visualise these spatio-temporal data in a variety of different ways, both static and interactive. But before exploring these, we will create a matplotlib colormap to display the 9 land cover classes with intuitive colors.

Now we can use the colormap with matplotlib to plot a static 2D map for a given time point.

Similarly, we could make a static plot showing maps for each point in time to visually examine change.

We can use holoviews to create an interactive plot - you should be able to use the slider below to see change through time in the same space.

3. Analysing Aggregate Change¶

In this section we’re going to examine land cover change across the study area. Our aim is to produce a bar plot that shows increases or decreases in land area for each land cover class between two points in time.

We can quickly identify the unique values in our land cover maps and how many times they are observed (i.e. how many pixels fall in each land cover class). We can use numpy’s .unique function to count the occurrence of each value, then output as a Pandas DataFrame.

It might be more intuitve (and useful) to work in units of area (e.g. square km) rather than pixel counts, so let’s convert the pixel counts to area (sq km):

We can see that some land cover classes are not present in any year (indicated by NaN). We’ll drop these and also reshape the data from ‘wide’ to ‘long’ for plotting.

From the Pandas DataFrame we can create both static and interactive bar plots. Here’s the static version with matplotlib.

We could also show this by calculating and plotting the change between first and final time snapshot.

4. Analysing Pixel-by-Pixel Change¶

In the last section we saw overall change in the land covers (e.g. decrease in Trees and Rangeland, increase in Crops). But which classes were changing to which other classes?

To understand the transitions between each class we need to analyse change for each pixel individually. We’ll enumerate individual pixels’ transitions for the first and last time snapshots using the a confusion matrix (using a function from sklearn and then visualise using a Sankey diagram (using holoviews).

First, the confusion matrix.

From the docs, the result is a:

confusion matrix whose i-th row and j-th column entry indicates the number of samples with true label being i-th class and predicted label being j-th class.

So for us, this means 2017 classes are the rows, with 2021 classes in columns (there are no predictions here - the 2021 observations are in place of any prediction). The number of cells that did not change are on the diagonal.

Now we can use the confusion matrix (after some manipulation) to create an interactive Sankey diagram using holoviews to visualise the transitions between each of the classes.

Using the data objects we created to make the Sankey diagram, we can also view the most frequent transitions in a table:

5. Analysing Zonal Change¶

Above we have analysed the data for our entire study area and on a pixel-by-pixel basis. But sometimes we want to work at a level between the entire study area and individuals, using zones that belong to some kind of vector geography. This is where raster zonal statistics come in:

Zonal Statistics uses groupings to calculate statistics for specified zones.

The statistics are summary statistics (e.g. mean, median, standard deviation) of all the pixels that fall in each zone. The zones might be watersheds, biomes or ecoregions (physically) or states, countries, counties or census districts (socio-economically).

Here, we’ll use municipality boundaries as zones. For Brazil these are freely available online in (zipped) shapefile format which we can read using GeoPandas (which inherits from Pandas).

Data Manipulation¶

Before we can use these data with the raster data we need to do some manipulation, including:

• setting the dtype of the CD_MUN series appropriately
• setting the map projection to be consistent with that for the raster land cover data (we’ll also do this for the study area polygon we made above)

And now we can plot to check things are aligning nicely and to visualise our zones (municipalities).

Let’s also make a static plot with matplotlib to visualise municipality boundaries over the raster land cover data.

Before we do our zonal statistics, to make comparison fair we will consider only those municipalities whose boundaries are completely within the study area.

First, remove municipalities from the vector shapefile that are not completely within the study area.

Next, create a raster from the new vector shapefile to mask cells from our land cover raster.

Let’s visualise the raster we’ve created from the vector shapefile of municipality boundaries.

And finally, mask the existing land cover xarray DataSet with this new xarray DataArray.

Now let’s check the mask worked by plotting the new (masked) land cover data with the municipality polygons overlaid.

One last piece of data manipulation is needed as the zonal_stats function we will use from the rasterstats package expects a 2D array in a format that can be read by the rasterio package (e.g. a numpy array). Hence data are for a single year of our data set.

Zonal Stats¶

Now we’re ready to actually calculate the statistics from the land cover raster by municipality shapefile polygon. When specifying that our input raster is categorical (as is the case for our land cover data), the zonal_stats function returns a list of dictionaries, one for each municipality polgyon. In turn, each dictionary provides cell counts (values) for each (land cover) class (keys).

We could examine these total counts, but in many ways it is more useful to consider the proportion of a muncipality that is covered by each land cover.

We could create a-spatial plots (e.g. bar plots) directly from this DataFrame. To create spatial plots (maps) we need to attach the spatial information for each municipality to create a GeoPandas GeoDataFrame:

And now we can do a static plot to visualise the proportion of two land covers across the municipalities.

An interactive plot enables linked pan/zoom of the two variables (although without consistent classification of polygons).

To examine change through time, we need to calculate the zonal stats for a second year, then calculate proportional change between the two time points. First the zonal stats.

Now calculate the proportional change between the two years.

From the summary statistics we can already begin to see differences in the trajectories of change. To map change spatially we finally merge with the spatial information.

Given that proportional change can be positive or negative, we should use a diverging colour palette.

This allows us to see that all municipalities decreased their Rangeland cover and increased their Cropland cover between 2017 and 2021, and to locate municipalities with extremes of change (which we could then go and further investigate).

Summary¶

This notebook introduce manipulation and exploratory analysis of classified land cover data. Specifically, it:

• used pystac-client to access information from the Microsoft’s Planetary Computer API and used that with odc-stac to read land cover raster data from an online repository.
• used matplotlib and holoviews to visualise the spatio-temporal land cover data statically and interactively.
• used from sklearn to create a matrix of pixel-by-pixel change and then holoviews to visualise that change using a Sankey diagram.
• used rasterstats to analyse zonal change with tabular data and visualisations.

Citing this Notebook¶

Please see CITATION.cff for the full citation information. The citation file can be exported to APA or BibTex formats (learn more here).

Additional information¶

Review: This notebook has been reviewed by one or more members of the Environmental Data Science book community. The open review is available here.

Dataset: 10m Annual Land Use Land Cover (9-class) from Impact Observatory for 2017-2021.

License: The code in this notebook is licensed under the MIT License. The Environmental Data Science book is licensed under the Creative Commons by Attribution 4.0 license. See further details here.

Contact: If you have any suggestion or report an issue with this notebook, feel free to create an issue or send a direct message to environmental.ds.book@gmail.com.