Discussion

Operationalised advantages of satellite images

Satellites have been used for many years to monitor changes, since the launch of the first LandSat satellite in 1984 developments have continued and now make available a vast amount of useable information . Many websites cater for quick access to processed data for various purposes, such as the aquamonitor, the forest watch on global scale (refs? Including the websites). However, the tool presented here is the first to be actually implemented for the Dutch River Authority (RWS) and to be used in the regulatory daily operational working process of river management. As far as we know this is also the first open access tool to present daily and automatically updating (Sentinel-2) satellite image analysis for floodplain management to a large group of end-users.

Performance and implementation

The accuracy of the tool is in line with earlier attempts for classifying floodplain vegetation in a similar number of classes and of the ‘traditional’ ecotope maps . The random forest classifier handles mixed classes such as ‘fields & grass’ unexpectedly well proving its robustness . Incorporating more images in the classification improved the result as seen in the overall accuracy of year-maps and day-maps. Clearly, classes that are separable on difference trough time benefit from a temporal data. For example, fields are the type of class that has a specific seasonal signature being bare at the start of the growing season and harvested as dissimilar stages and dates.
We see some opportunities to improve the classification for year-maps and single-date maps. Separation of classes in year-maps showing a seasonal signal could be improved when incorporating pixel-level fitted parameters on sinusoidal function of yearly NDVI as bands . For single date (image) maps and year-maps, the inclusion of SAR data could provide some extra resolution . However, experiments with ‘raw’ Sentinel-1 data (hh, hv; data not shown) did not show significant improvement (order of about 1-2% to total accuracy). Another approach could be a combination of segmentation in super-pixels and the use of texture within those super-pixels . This can also be a way to incorporate SAR data into the classification . However, for single date maps the classification speed determines the user-experience greatly. Therefore, there is a speed trade-off for incorporating new procedures and data in on-the-fly classification. The pre-classified year-maps do not have that disadvantage.
The accuracy was one of the most important topics in the discussion with the end-users. Beforehand we could estimate the accuracy based on quick scans and existing knowledge, but these numbers did not mean much to the end-users. During the first development year results were quickly shared and maps were used in the field. This experience helped translate the statistical accuracy in a form of trust in the produced maps. The vegetation-monitoring team could use the map in discussions with stakeholders and have no doubt about the map contents.

Development process

Developing a new monitoring tool for operational floodplain management is a process that is highly interactive and a strong dialogue with end-users is needed to create a useable tool. Features like GPS location indication for field use and downloadable images were implemented in a second iteration of the tool. Also, as the end-users started to download small AOIs (as in on-screen views) and mosaicking these in GIS, the download option evolved from downloads of single user selected classifications of the AOI in the viewer to downloads for total covered area. We further added the year-maps after different users asked for a “standard year map” to be able to refer to in reports and other stakeholders in order to standardize the yearly assessment of the vegetation state. It is foreseen that there will be additional changes to the tool out of the continued use and experience in the coming period. Also, a new or updated fixed map-layers will need to be included when made available. For instance, a new legal vegetation map will become available in 2020.
The process of creating this tool has taken several years of intense discussions between researchers and managers at the national water board. The interplay between scientific and technical advances, and management requirements was helped by labelling this project as an innovative project within the framework of cooperation. The end-user interaction to come to the here presented tool must not be underestimated in the creation of an accepted and useful tool. The availability of Google Earth Engine with its continued up-to-date image library greatly speeded up and facilitated the prototyping and operationalizing of the project.

Reflection on use and other applications

The vegetation monitor is primarily designed as a quick and easy screening tool to identify those areas in the Dutch floodplains that in need of maintenance and is always used as a starting point for the dialogue with the responsible land owners and must be complimented with field visits for the final check on the correctness of the initial assessment. Next to this there are many potential possibilities to use the tool also in a wider context, for example to assess the impact of the changing lay-outs of vegetation distribution on flood water levels when these maps are directly used as input for the flood risk models. Also, the tool can easily be adapted to be useful in other river systems, providing that some ground truth data is available as test and training data. Several other initiatives are currently undergoing to classify the vegetation of the floodplains, but most often this is done in a GIS environment, making it less accessible to the general audience. Especially in large scale data-poor areas the technique may give a first order estimation of the current status and allow also for analysis on changes over time.

Recommendations

During the creation of the vegetation monitor there was discussion regarding the absolute accuracy of the individual classification results and which percentage of accuracy was deemed acceptable for the evaluation of the status of the vegetation. It took time to acknowledge that reaching 95% accuracy was not only technically not feasible, but also not necessary to give an overall judgement of the status of the vegetation. The annual maps with an average accuracy of between 80 % and 86 % are now being used and give a sufficient first indication of the situation. The communication that this tool is a first screening and that field visits will remain necessary in the final judgement has helped to accept that a certain level of inaccuracy is unavoidable.
During the stakeholder process to define the terms of reference, it was clear that different types of users have different requirements. The stakeholder process is therefore very important in defining the final features of such tool.
The time series analysis feature has added benefit in judging a single classification, as it aids in checking for the consistency of the classification and potential fluctuations and dynamics through time, e.g. in relation to interannual fluctuations in river discharge and weather patterns and the response of vegetation to this.