Continuing our series on the use of RapidEye data in geospatial projects, for this edition of The Geospatial Times we demonstrate the value of RapidEye satellite imagery for automated extraction of rivers and streams across large areas. These features can be particularly challenging to extract manually given the complexity of riparian zones, so finding a reliable technique to do this in an automated fashion can be a huge time (and cost!) saver.
The RapidEye Constellation
RapidEye is a constellation of five 5-meter medium resolution satellites each offering five spectral bands of information. The RapidEye constellation offers a daily revisit time to every location on the planet with a huge footprint that is 77-km wide. The data is priced extremely competitively with a base price of $1.28 per square kilometer for all five spectral bands – academics receive a discount on this price. RapidEye adds a fifth band, the red edge, to the ‘traditional’ multispectral set of blue, green, red and near-infrared (NIR). The additional spectral data available in the red edge band allows users to extract more useful land ‘information’ than can be extracted from traditional 4-band imagery sources.
River Channel Extraction
While RapidEye has lower resolution than data offered by high-res satellites, its wide area coverage and five spectral bands make it an ideal solution for automated extraction of stream and rivers over a large area. To complete this analysis, we used ENVI 4.8 and extracted rivers from a scene of RapidEye data collected over the San Luis Valley, Colorado on June 19, 2011 utilizing the built-in Supervised Classification algorithms.
In the first step of this automated extraction, we built a training set of river channels utilizing the Region of Interest (ROI) tool in ENVI 4.8. This training set contained seven polygons (or 318 pixels) over the center waterways of large to medium sized rivers evenly dispersed throughout the RapidEye scene. Next, we ran a Parallelepiped Supervised Classification filter set to identify pixels in the imagery that were within three standard deviations of the mean spectral values of the ROI.
An inspection of the results revealed that we had a lot of room for improvement as large stretches of smaller rivers were missed. As such, we added in two more polygons to the ROI over smaller streams for a total of 9 areas and 345 pixels. The same Parallelepiped filter was run again on the entire RapidEye scene yielding the results you see below labeled, ‘Iteration 2’.
An inspection of the results from Iteration 2 showed that we could still do a better job finding the smaller streams; so we added in 2 more polygons (for 394 total pixels) over streams that were missed in the supervised classification. The same Parallelepiped filter was run using the third version of the ROI – and again the results were disappointing. This time we had many false identifications of healthy vegetation and bare soil which suggests that the polygons we added to the ROI contained these surface features.
For the fourth and final iteration, we removed the two polygons added to the ROI in Iteration 3 and replaced them with smaller areas. The two new polygons we added were over small streams, but this time we focused on pixels that appeared very dark suggesting they were ‘pure’ water – our new ROI contained 11 polygons covering 353 pixels. The results of Iteration 4 were the best we achieved but there were still issues to be found. In the animation below, we circled one area of concern with an orange ellipse. You can see that this field was ID-ed as a river in all four iterations suggesting that: (1) it was watered right before this imagery was collected; and that (2) the dominant spectra detected by RapidEye was that of water rather than the healthy vegetation it saturated.
The results of this analysis suggest that medium resolution RapidEye imagery is a useful tool for the automated extraction of river channels. In our analysis, we identified 191,794 pixels of water or freshly watered areas which cover more than 1,180 acres! While there is certainly room for improvement with the analysis put forth here, the technique is obviously sensitive as it detected freshly watered agricultural fields. Taking this logic farther, one can also theorize that the streams missed in this analysis had little to no water flowing through them or were dominated by emergent vegetation; and that utilizing imagery collected during the spring with higher flow rates should help to maximize detection rates. Another idea would be tweaking the standard deviation threshold so that identified ‘river’ pixels would have to match the mean spectral values of the ROI more closely. As a final note, it is much easier to manually delete misidentified pixels than it is to hand-digitize a complex river channel.
If you would like to find out more about using RapidEye for your academic, engineering or any landscape study, let us know at email@example.com or (303) 993-3863.
Leave a Reply