Vegetation Detection
Vegetation Detection
Given the importance of macrophyte cover in understanding performance of constructed wetlands, wouldn’t it be great to have a simple and cost-effective way to undertake quantitative estimates of plant cover, and to track changes in cover over time? This page outlines a simple method for estimation of emergent vegetation cover using the NDVI calculation from readily available hi-resolution aerial imagery that includes the Near-Infrared (NIR) band.
Remote sensing techniques are rapidly changing the way environmental data can be captured over large spatial areas. Application of these methods to remote sensing of vegetation cover and/or land-type classification generally rely on satellite imagery with a large number of spectral bands, along with calibrated values of surface reflectance for each of these bands. However, this satellite imagery is often at resolutions too large for a WSUD asset-scale analysis (pixel size 10m-50m+) with aquisition of the required high-resolution datasets (where available) generally being cost prohibitive for many asset owners.
Given that some providers of aerial imagery do include capture of the Near-Infrared (NIR) band within more affordable subscriptions, we decided we would see if this data could be used to gain an understanding of vegetation cover in wetlands, and to see how this cover may change over time.
The images from 23/10/2019 shown above were sourced through our MetroMap subscription for a constructed wetland system located in Lara, Victoria. The left image shows a conventional R-G-B image of the site, with the image on the right showing the bands NIR-G-R for the same date. Images were downloaded to cover the entire system which is made up of 4 large wetland cells, 13 smaller sediment basins, and 2 planted (transfer) swales. Aerial Images from were also downloaded from 12/10/2018 to enable comparison over the two years. The wavelengths of these respective bands are shown in the figure below (image credit: MetroMap).
Healthy vegetation reflects large amounts of NIR, making these images useful to understand plant cover and plant health. By extracting the NIR and red bands from the image, we will be able to perform a calculation of the Normalised Difference Vegetation Index (NDVI). This calculation has been widely used for vegetation detection and analysis and is explained in the following video.
One concern for this method arises in that our MetroMap image is in a highly compressed lossy jpeg format. A quick search of the online blogs by those in-the-know indicates that using these images is not as accurate as many other methods. This is because this format uses just a digital number (DN) with a value between 0-255 for each pixel in each band, which may not correlate directly with the actual reflectance value from the raw image. However, given (i) some of the other limitations associated with satellite imagery discussed above, (ii) that the NDVI calculation is normalised by its own values, and (iii) that we are mainly trying to only detect where vegetation is present (ie. a yes or no) we thought it would still be worth giving it a shot. The trade-off in image compression is the significantly better resolution of the image when compared to the satellite data, with some discussion of compression effects of classification discussed in more detail in studies such as this one.
To help mitigate some of the effects of this jpeg compression as well as from different lighting/weather conditions between the photo dates, image bands from each year were normalised to push the DN values out to the full range of 0-255. Together with the fact the NDVI formula is also self-normalising, this should help see a robust and repeatable result that is able to (at least somewhat) deal with changes in lighting from different photo dates that always make these types of analysis difficult regardless of the source of the image. The NDVI results are shown in the following image.
NDVI will always return values of between -1.0 and 1.0. Interpretation of exact NDVI values is shown to vary depending on who is providing the advice, but generally areas of no vegetation or completely dead vegetation will return a value of 0 or less. Areas of water should also return a value of 0 or less (as discussed in the video). Values of 0.25 or higher can often be considered to indicate (at least moderately) healthy plants, with values between 0 and 0.25 perhaps being the hardest to classify as being vegetation or not. This question can be compounded by the fact that many wetland species are dormant during winter months, and may have foliage which tends to be much closer to yellow than green during those seasons. To take all this into account we decided to classify our NDVI scores as follows:
Polygons were then created for areas of vegetation classification based on the above NDVI values. These were then clipped to the waterbody polygon extents (shown in blue). The results were, well… they appeared to be very good (see below).
The NDVI based classification appeared to do a good job at highlighting the areas of emergent vegetation throughout the waterbody extents as well as removing all paths and open water zones. We do note that the MetroMap NIR image indicates that there were some small additional areas in the centre of the wetland where aquatic vegetation is present (but below the waterline), and given the turbidity of water at the time of the photo that this is not being detected using this method. (NDVI calculations in these locations returned values between -0.15 and 0.05). This method was applied to the imagery from both Oct 2018 and Oct 2019. The following figure provides the vegetation cover results on an asset by asset basis, as well as how much change was observed over the 12month period. Feel free to click around the map provided to explore the data for yourself! (tip: click the arrows on the top left of the map to explore different layers)
In summary it appears that this method can provide a cost-effective and repeatable method for calculating emergent plant cover across very large assets, or many assets spread over a large area. Although the use of compressed imagery may mean that the classifications should not be used as absolute delineation of the plant extents, it is likely this method could be used to quickly estimate plant cover and observe trends in plant cover over time using a simple formula rather than ‘black-box’ machine learning classifications. Further optimizations could likely be achieved by different approaches of normalising the jpeg data, as well as further trial and error of the minimum NDVI value used to indicate that vegetation is present.
This data could be used to help inform where more detailed site based assessments may be required, and help answer questions like “where is wetland plant cover very high or very low?”, “where is plant cover in sediment basins too high?”, or “which sites are seeing the biggest gains or losses in vegetation over time?”. This method has some additional considerations which would also apply to other methods of remote sensing that include: the need to understand water levels and seasonality of plant species at the time of imaging and also checking if tree canopies or algal blooms are influencing the results. It can also not indicate where the vegetation cover is actually nuisance species rather than desirable ones, which is best done via ground truthing alongside this sort of analysis.
Next step: to capture some super high resolution multispectral imagery via drone survey, and compare those results…
This information has been prepared solely for demonstration purposes, and is issued in confidence for the purposes only for which it is supplied. Non-authorised use of this information in any form whatsoever is prohibited. No liability is accepted with respect to any use by any other person or organisation. Whilst all possible care has been taken to ensure the accuracy of information provided and the relevance and appropriateness of the information, the authors accept no liability or responsibility relating to any action or lack of action taken based on this information and/or any associated work.