Land cover / forest mapping using Sentinel-1 data

Land use and land cover are most often mapped with optical data like the data from the Sentinel-2 satellites. In highly cloud-plagued areas like e.g. in the Tropics or Arctic regions, radar data from Sentinel-1 can also be used to make land cover maps if the land cover types are general. Variation of backscatter over time is different in various land cover types. For instance, mature forest has a very stable backscatter level while many agricultural areas show distinct variations depending on the crop type and crop calendar. When combining acquisitions from different parts of the year, these variations can be used to map land cover types. VV-polarized Sentinel-1 data are used because this is the most widely acquired data type globally (present in both dual and single-polarized scenes). The algorithm chosen for land cover mapping in the F-TEP land cover service is the random forest algorithm as implemented in the Orfeo toolbox. Based on samples of land cover classes, a set of decision trees is produced. For each tree, a random set of spectral bands at each decision node is chosen. The final land cover class is then chosen based on majority voting between the output of each decision tree. The benefits of the random-forest algorithm include good adaptation to many different types of classification problems as well as independence of assumptions on statistical distribution of band-wise backscatter values within a (land-cover) class. The latter property is important for radar data since the distribution of radar backscatter data tends to differ from normal distribution.

Sentinel-1 Land Cover service data inputs

The Sentinel-1 Land Cover service within F-TEP requires two data inputs:

  • A shapefile containing reference data for training the classifier
  • A time series of Sentinel-1 VV-polarised (SVV) images or a collection of Sentinel-1 VV-polarised images collected in a Databasket

The reference data shape file must include polygons, each polygon covering area belonging to just one land cover class. For each target class, the reference data should include several hundred pixels in the used input data at the selected pixel size. The reference data shape file – and all supporting files like the .dbf file, .prj file and the .shx file - must be packaged into a single .zip file. The zip file must be uploaded to F-TEP using the Files feature in the Manage/Share section, or alternatively placed in a web location that is accessible via a public URL (Universal Resource Locator). The Sentinel-2 Land Cover service tutorial shows a way to produce an acceptable shape file using GoogleEarth.

Running the Sentinel-1 Land Cover service

The Sentinel-1 land cover service requires defining parameters in the Service dialogue panel ("workspace") shown below:


  1. Select the Input Data:The input Sentinel-1 images (GRDH products) are best collected into a databasket, which is then dragged to Input Field 1. In case of just one S-1 input data, you can drag that from the search results window to Input Field 1. Selection of satellite image data is covered in a separate tutorial.
  2. Define the Reference data shapefile: Here you need to define the reference data shapefile to be used, i.e. a .zip package file containing all the files belonging to the shapefile.
    • If you have previously uploaded your own reference data to the platform (using the Files feature in the Manage/Share section), do a search for reference data to locate your file. After identifying your reference data file in the search results, click the Workspace icon on the left to return to the service definition so that your previous inputs are intact. Drag your reference data file from the search results view to Input Field 2.
    • Alternatively, you can write in Input Field 2 the URL of a reference data (zip packaged shapefile) that you have available in a publicly available web location. An example of a valid reference data address is
  3. Define the field in the Reference data shapefile that will be used to train the classifer: Write the name of the class code column (“tessellate” in this example) of your reference data shape file to Input Field 3.
  4. Target CRS identifier: Write the code for the output Coordinate Reference System (CRS) to Input Field 4.  This must be a string beginning with “EPSG:”, and the following numeric code must be a valid system as defined by the European Petroleum Survey Group.  For UTM zones in northern hemisphere and with the WGS84 datum, the numeric code is 32600 + the UTM zone number.  An example of a valid CRS specification is: EPSG:32615
  5. Area of interest (AOI): Define the AOI to be processed in Input Field 5. This can be copied from the area drawn on the map by pressing the Copy from Map button. Alternatively this can be added as simple text, although it needs to follow the WKT POLYGON specification. An example of a valid AOI specification is: POLYGON((-92.906633 16.190411,-92.066559 16.188383,-92.070266 15.376645,-92.907004 15.378567,-92.906633 16.190411)). The AOI should be large enough that it includes enough reference data for all target class.
  6. Define the DEM to use. A DEM is needed for ortho-rectification and terrain correction of Sentinel-1 radar images.  The identification string for the SNAP software is typed to Input Field 6.  The default is “SRTM 1Sec HGT”, which is valid between -60 and 60 latitudes.   
  7. Define the Target Image Resolution. Write the requested output image pixel spacing (in metres) to Input Field 7.  For Sentinel-1 radar data, feasible values are between 20 and 100 m. Example: '20'

After all input fields have been filled in, the Sentinel-1 land cover service is launched by clicking the round play button at the bottom right corner of the input dialogue area. The resulting output can be downloaded, re-used or displayed in one of the GUI applications such as Sentinel Toolbox or QGIS. Tutorials on opening data within the Sentinel Toolbox and QGIS are provided. The image below shows a generated (with 4 images covering the AOI variably) land cover map within QGIS. A longer time-series covering the whole area of interest is needed for a good land cover map.