| dc.description.abstract | In this MSc research project, the feasibility of automatically deriving habitat maps from UAV derived orthomosaics using deep learning is explored. Habitat maps are crucial for preserving biodiversity in nature areas, and traditional methods of creating them are costly and time-consuming. Creating habitat maps typically involves high-quality fieldwork, which is expensive and leads to infrequent updates, even though dunes can be highly dynamic. This project focuses on the Dutch National Park Zuid-Kennemerland and addresses the problem of automating the creation of habitat maps by exploring the use of deep learning algorithms and increasing the performance by using hyperparameter tuning. In this research we show that using an annotation tool called doodler and a ResUnet with class weights from segmentation gym performs the most optimal. The habitats demonstrates that the ResUnet model outperforms the segformer and UNet models in classifying sand, grey dunes, white dunes, and shrubs in coastal dune images. However, there are still difficulties in differentiating between grey dunes and white dunes. The use of class weights during training improves overall predictions and reduces tiling issues in the fully predicted image. These findings contribute to the understanding of effective image segmentation techniques for habitat mapping in coastal dune environments in the Netherlands, emphasizing the importance of model selection and the use of class weights for improved performance. We anticipate this thesis to be a starting point in the automation of habitat mapping for coastal dunes. Some hyperparameters like image augmentation can still be tested. Also a popular trend in deep learning for images is transfer learning which might achieve even better results. Furthermore accurately automating this process can save considerate time for researchers who create habitat maps for coastal dunes and makes the tracking of biodiversity within these areas more consistent. | |
