Samenvatting

Within the Province of Zeeland, tree removal is generally allowed, but it is required to notify the province and replant an equivalent number of trees within three years. However, there is limited insight into tree removal that is not reported through this mandatory notification system, resulting in unauthorized tree removal remaining undetected for longer periods of time. This study investigates the feasibility of using geospatial data such as satellite imagery, to support in the detection of potential violations of tree removal.

Multiple approaches were explored and compared, including NDVI-based change thresholds, deep learning models and a semantic tree crown segmentation approach. While deep learning methods such as CNNs and Siamese networks did not achieve satisfactory results, the best performance was achieved using a semantic tree crown segmentation model. In this approach, tree loss was quantified by directly measuring decreases in the detected tree crowns over time, reaching a recall and specificity of \~78%. Furthermore, this same approach can be used to detect replanted trees by measuring increases in vegetation between two points in time, allowing for the monitoring of replanting of the removed trees.

While the method does not fully meet the predefined success criteria of 80% for both recall and specificity, it shows potential as decision-support tool for the monitoring of unauthorized tree removal. It can be used as practical tool to help the Province of Zeeland find areas where tree removal and replanting is more likely to occur, improving the efficiency of manual inspections and support more targeted searching of the area.