Posted on Nov 03, 2017 by Anisha
OpenStreetMap (OSM) data is a fantastic resource which provides free to use mapping, worldwide. Increasingly, it is providing excellent street level mapping even in remote areas. However, being user-generated, there can be a degree of uncertainty as to its accuracy.
One way to decrease this risk is to use up-to-date satellite imagery to verify the OpenStreetMap data. Here is a comparison of the current OSM with a Sentinel-2 dataset displayed in natural colour:
Figure 1: OSM compared to Sentinel-2 data displayed as RGB (Bands 4,3,2)
At first glance the OSM data looks pretty good, but looking more closely reveals some inconsistencies.
Figure 2: Detail showing variation between OSM and Sentinel-2
Examining the data by eye does give a clear indication as to where change has occurred but over a large area this could be time-consuming. It would be better if we could start to automate this process.
If, for instance, we wished to compare the residential areas in OSM with the recent Sentinel data, we can use the classification functions within ERDAS IMAGINE to automatically extract these areas. Traditional spectral classification tools are not really appropriate if one is looking to extract just one landcover type. However, we can make use of the Spectral Analysis Workstation to do just this.
Figure 3: The Spectral Analysis Workstation
This produces what is in effect a probability layer where the higher the value, the more likely it is that that pixel is a match to the target spectrum.
Figure 4: Comparison between Sentinel imagery and Spectral Analysis Workstation output. Brighter areas are more likely to be built up.
As with any thematic dataset, this can be thresholded to extract just the areas we are interested in; in this case the residential areas. We can use Spatial Modeler to do this:
Figure 5: A simple threshold model in Spatial Modeler
The output is a little speckled, so we can use some cleaning tools to produce a more vector friendly output:
Figure 6: A more complex model that tidies the raster and outputs a shapefile. Note that much more of the residential area has been captured compared to the OSM data.
On the left we can see the current OSM residential areas in blue compared with the area generated from Sentinel-2.
In a similar way, here is an example where the wetland class of OSM (in blue) is not a very good match to the conditions on the ground. This may be due to poor user mapping, or perhaps seasonal variation. In any event, it can be seen that Sentinel-2 provides a more accurate base layer so that wetlands (seen here in dark reds and greens) can be accurately mapped for cross country routing, for instance.
Figure 7: Wetland area captured from Sentinel-2 data. Note: this is much more accurate and complete for this season than the OSM data.
Of course, there are times when one cannot obtain up-to-date optical imagery due to cloud cover. Even though it’s not quite so clear, one can still extract features such as wetland from Sentinel-1 SAR data:
Figure 8: Wetland highlighted on Sentinel-1 SAR data.
Note: The Sentinel-1 and -2 data were obtained at different times so the area of wetland is slightly different.
Not all landcover features are amenable for extraction/interpretation at this spatial resolution. However, Sentinel data can provide a quick, easy and up-to-date method for verifying and/or editing OSM data.