Congratulations to Håkon Næss Sandum at NMBU on defending his PhD with the thesis title,
Towards operational deep learning-based forest stand delineation!
🌲 How can deep learning support forest mapping in practice?
Dividing forests into stands provides a foundation for forest inventories and management planning. Doing this manually, however, is time-consuming and relies on individual judgement. Håkon’s research explored how AI can learn from expert-drawn maps, combining aerial imagery with canopy height data to map forest development stages.
Key findings:
- Deep learning models could closely approximate expert mapping of forest development stages, with errors mainly occurring between neighbouring stages.
- Canopy height data derived from aerial photographs performed comparably to laser-scanning-based data, offering a practical alternative captured at the same time as the imagery.
- Updated older stand maps could support model training despite inaccuracies in their labels—reducing reliance on costly new manual annotations.
The practical takeaway: a promising tool for generating preliminary forest maps for expert review, supporting more efficient inventory workflows. The approach does not yet separate neighbouring stands at the same development stage, but provides a stepping stone towards delineating individual stands.
A big thank you to the Evaluation committee, with the opponents Jari Vauhkonen, University of Eastern Finland, and Ruben Valbuena, Swedish University of Agricultural Sciences (SLU), the Committee coordinator Line Nybakken and the PhD supervisors Terje Gobakken, Hans Ole Ørka and Oliver Tomic, all NMBU!
Congratulations, Håkon, on this milestone and your contribution to more automated forest mapping!
You can read the already published articles of his thesis here: