Mapping the tree species distribution of Rumperöd forest using UAV-RGB imagery and object-based classification

Forest ecosystems are important actors in climate change mitigation strategies through their abilities to act as both carbon sinks and sources. The commonly used forest management strategy of clearcutting has been suggested to be paired with large greenhouse gas emissions, increasing the need for research on the climate effect from alternative management strategies. Rumperöd forest is managed through selective thinning, which is a less common strategy in Sweden, maintaining a continuous forest cover through smaller harvests while promoting favorable stand functions and tree species diversity. Ecosystem carbon exchange is monitored with a flux tower in Rumperöd, creating opportunities of quantification of the effects of selective thinning. To enable within-stand analysis of carbon fluxes and management effects, mapping within-stand differences is essential. This thesis explores the method of species distribution mapping utilizing remote sensing techniques and drone-captured RGB images (UAV-RGB imagery). Mapping is done on an object-basis, first segmenting each tree crown using the Tree Segmentation Model and then classifying each segment with a Random Forest classifier. Segmentation achieved 67% similarity to manually segmented tree crowns, containing several types of errors. Classification was done with spectral indices as model inputs achieving 71% accuracy compared to ground collected data. The accuracy was lower than found for similar studies, however the distribution map manages to show known large-scale distributions of the forest and to separate an increased number of species than earlier attempt at the same site. This thesis shows that UAV-RGB imagery can be used for species distribution mapping in complex forests with varying accuracy between species, suggesting multiple ways for future method improvements.