Classification of Coastal Dune Vegetation from Aerial Imagery with a Convolutional Neural Network
Summary
The Dutch coastal dunes house a rich and dynamic ecosystem with a large variation in habitats. However, they are currently threatened by geomorphological stabilization and subsequent encroachment of shrubs and tall grasses, which overtake valuable grey dune grassland. To understand such vegetation dynamics at larger scales, vegetation maps are key. While traditionally, constructing these maps from field surveys is very time- and cost-intensive, an alternative approach can be taken by training a machine learning model to classify vegetation from remote sensing data. A particularly promising model is the Convolutional Neural Network (CNN), a type of deep learning model specifically designed for pattern recognition, which has been proven highly accurate in mapping vegetation in various ecosystems. However, no vegetation mapping CNNs have been developed for coastal dune ecosystems yet.
In this thesis I present a CNN capable of mapping coastal dune vegetation on a plant community scale with an overall accuracy of 75%. The CNN has a U-net architecture and is trained for RGB orthophoto tiles with spatial dimensions of 10 by 10 m. First, a vegetation survey of the study site, the blowout complex north of Bloemendaal, was conducted to better understand its vegetation. Additionally, an Uncrewed Aerial Vehicle survey to obtain an orthophoto and additional elevation data of the study site was conducted. The CNN was trained using 112,500 m2 of manually classified vegetation maps as reference data. It was found that CNN performance per class is positively related to the surface area of that class in the reference data. Furthermore, increasing the input tile size was shown to increase overall CNN accuracy. Conversely, including elevation data as additional input information was not found to make a significant difference in CNN performance. These findings can contribute to improving the CNN further in the future.
The results of this thesis demonstrate that it is possible to use a CNN for mapping coastal dune vegetation, obtaining a vegetation map that is ecologically relevant and can be compared to other vegetation studies that use Natura 2000 habitat types. Although this CNN must still be optimized further, it could potentially be used for long-time and large-scale vegetation mapping, which could greatly improve our understanding of vegetation dynamics in coastal dune ecosystems.