Gertjan Geerlinga,d, Ellis Penninga, Gennadii Donchytsa, Stanford Wilson b, Joshua Iked, Rik van Neerd,
Rick Kuggelijnb
aDeltares, Boussinesqweg 1, 2629 HV, Delft, the Netherlands
bRijkswaterstaat WVL, Zuiderwagenplein 2, 8224 AD Lelystad, the Netherlands
cRijkswaterstaat Eusebiusbuitensingel 66, 6828 HZ Arnhem
dInstitute for Science in Society, Radboud Universiteit, Nijmegen
Keywords — monitoring, floodplain vegetation, remote sensing, google earth engine Introduction
The National Water Authority (Rijkswaterstaat or RWS) has the task to ensure an efficient discharge of water (from the Rhine and Meuse) at high water levels. The floodplains and their vegetation partly determine the water levels at high discharges; therefore maintenance of the floodplain’s morphology and vegetation is essential.
The floodplain configuration at which a safe water level is guaranteed (at a designated water level) is mapped into the “vegetatie legger” and part of the legal framework for the watermanager called “waterwet”. The
‘vegetatie legger’ consists of a map of the floodplains in which the vegetation is
summarised into 6 classes. If the current state of the floodplains is different and the local hydraulic roughness exceeds that of the locally allowed vegetation type, vegetation
maintenance is needed.
To compare a current state of floodplain vegetation to the ‘vegetatie-legger’, RWS needs a map of the current state. The current mapping procedure (ecotope maps) is a proven and robust method, but the production of the ecotope maps has a processing time of one year and its mapping cycle is every 6 years. Six years is to long for timely vegetation management, optimally one would like the information from within the same year for the whole management area comprising of Meuse and Rhine.
System Idea and basic methods of the monitoring application
The initial idea was to create an application that gives insight in the current state of a selected floodplain or river section and
compares that state to the ‘vegetatie legger’. It was decided that the developed application should allow all landowners in the floodplain to be able to check their ‘vegetation status’, so a web-based interface was chosen. Sentinel-2 satellite images were chosen as principal data source because of the high temporal coverage (every 5 days), free availability and resolution (10x10m). Using Google Earth Engine (GEE, https://earthengine.google.com/) makes it possible to implement an on-the-fly
classification for areas of interest. The random forest classifier is trained using existing vegetation structure maps with about 200 points for each ‘vegetatie legger’ class.
Result
The vegetation monitor front-end is shown in Figure 1. You can visit
https://www.openearth.nl/vegetatiemonitor/ to test the application. The user can zoom to an area of interest, look at the original
vegetatie-legger and flow paths with highest flow, then select sentinel images by date, classify an selected image, and calculate the difference with the ‘vegetatie legger’.
Additionally, classified images can be
downloaded (geotiff). Per land owner polygon a PDF summary comparing the ‘vegetatie legger’ vegetation distribution (% surface area of types) with a classified image is made available for downloading. The back-end is rooted in python, MapBox and the maps and computations are performed in Google Earth Engine (GEE).
The classification results vary per area of interest and selected image. The best results were obtained using August or September images, when vegetation types are optimally
developed, and reach 71% total accuracy based on training and testing with vegetation structure maps from 2017. Field trials showed the maps provided sufficient detail to be able to recognise past vegetation management activities by end-users.
Figure 1. Screenshot of the web-interface of the
vegetation-monitor application. On the left side you see the various spatial layers, including “classificatie” which contains the classification. The right side shows the area of interest, in the screenshot a classified image is drawn on top of a true-color version of the satellite image. Application experiences
The vegetation monitoring tool is the first tool that gives up to date information of the current status of floodplain vegetation in the Rhine and Meuse floodplain areas. The revisiting time of sentinel is much higher than the annual aerial photograph survey (which is still used for verification). Comparing the ‘vegetatie legger’ with the current state allows an indication of how much “maintenance space” is available before interventions must be carried out (such as throwback of vegetation succession). Experiences of RWS show that the monitoring tool helps a lot in the discussion with nature organizations who are responsible for the maintenance of the vegetation in large parts of the floodplains. During the annual field
check-ups the tool was used on tablets during field visits to areas which were classified as being rougher than the vegetatie legger. There
were several reasons why spots were
classified rougher: during the hot summer, the river dried up which had consequences for the shore areas going from water to sand. Quick developments of willow storage along the shore or at dried up gullies were identified quickly and efficiently. Although the classification of the tool was not always
correct, given the fact that this is a first version; the reactions and feedback from RWS was very positive and the tool has been widely accepted by other end users such as Staatsbosbeheer, Natuurmonumenten, and other cultural and landscape foundations.
Outlook 2019
In 2019 the vegetation monitor will be further developed to improve the classification algorithms and explore if vegetation prediction can be added. The user experience will be enhanced by continued dialog between users and developers. Classification improvements under consideration are refining the training set by eliminating outliers in the training data and testing additional data such as newer LiDAR and Radar (Sentinel-1). A first version of a vegetation-prediction module was prototyped in 2018, but needs more ‘vegetation succession rules’ and validation using Landsat time-series for hind-casting to validate the vegetation succession rate as currently estimated by experts.
References
Geerling, G, Penning WE, Donchyts G, Wilson S (2018). Mapping and change detection of floodplain vegetation by remote sensing (Sentinel-2 in Google Earth Engine) for water management on river delta scale. Conference paper International Symposium on Ecohydraulics tokyo, August 2018.
Penning E en Geerling G (2018) Vegetatiemonitor- werking en algemene resultaten. Deltares rapport 11202298-001, Delft.