Mapping Downed Coarse Woody Debris using high resolution RGB UAV images and Convolutional Neural Networks | Maalahopuun kartoitus konvoluutionaalisten neuroverkkojen avulla korkearesoluutiosista UAV RGB kuvista

Dead wood is an integral part for forest biodiversity in boreal forests. 5000 (25 %) of Finland’s forest dwelling species depend on decaying dead wood during their life cycle. The loss of dead wood in forest ecosystems has been identified as the number one reason for species endangerment. Conventional dead wood mapping is done by counting and measuring dead wood from field plots or by aerial laser scanning, both of which can be timeand resource consuming. UAV-borne aerial imaging provides cost effective and high spatial and temporal resolution in comparison to conventional aerial imaging and satellite-based imagery. A convolutional neural network (CNN) is a deep learning algorithm that has shown promise in recognizing spatial patterns. The strengths of CNNs are end-to-end learning and transfer learning. CNNs have been used for mapping both standing and downed dead wood. This thesis aims to further investigate the usability of a method based on detecting downed coarse woody debris (CWD) in a coniferous boreal forest from RGB UAV-imagery using a CNN based segmentation approach. CWD was digitized from an orthomosaic created from UAV-imagery. CWD was digitized from 68 square shaped 100 x 100 m virtual plots surrounding 9 m radius circular field plots. The plots were divided into 57 training plots for training the CNN and 11 test plots for evaluating the CNN model performance. The effect of different loss functions and the effect of data augmentation on model segmentation performance was evaluated. The number of digitized and segmented CWD objects were compared to the number of CWD objects from the field plots and the effect of canopy cover and basal area on the detection rate was assessed. The CNN model segmented 324 m ² of CWD from the 11 virtual test plots, from which 469 m ² of CWD had been digitized, resulting in a 69 % segmented-to-digitized CWD ratio. The model with the best performance achieved a precision of 0.722, a recall of 0.500, a Dice-score of 0.591, and an intersection over union (IoU) of 0.42. The sample size of field measured CWD from the field plots was relatively small and neither canopy cover nor basal area was found to have a statistically significant (P = 0.05) effect on CWD detection rate. For the digitized CWD detection rate, canopy cover had a p-value of 0.059 and basal area a p-value of 0.764. For the model segmented CWD detection rate, the p-values were 0.052 and 0.884, respectively.