Browsing by Author "Nyman, Johannes"

Gathering information on forest structure is vital in estimating forest biodiversity, carbon stocks and temporal changes in standing forests. Currently the only viable method of collecting such information in vast areas is remote sensing (RS). Two commonly used RS methods for acquiring high resolution three dimensional information on forest structure are airborne laser scanning (ALS) and digital aerial photogrammetry (DAP). In quantifying forest structure, the distributions of tree basal areas have been used because the variation in tree sizes is closely linked to the whole concept of forest structure. Retrieving information on these distributions can be done by modelling the relationship of in situ measured distribution indices and the remotely sensed elevation information. One of these distribution indices is the Gini coefficient which has been shown to be a prominent index in describing the forest structure from ALS data. In this study, DAP data was gathered with an unmanned aerial system (UAS) from the vicinity of the Lammi research station with the intention of investigating its suitability on modelling forest structure by using Gini coefficient (GC). Airborne laser scanning data retrieved from the National land survey was used as a comparison dataset. The in situ measured field data consisted of tree circumference measurements from 50 circular plots (r = 5m). From these measurements, the tree basal areas were calculated and the plot level Gini coefficients determined. A comprehensive set of plot level point cloud variables were also calculated from both ALS and DAP point clouds. The most important predictor variables were chosen from the point cloud variables with an automatic exhaustive variable selection function. Then, beta regression modelling was applied to both sets of predictor variables and the best GC models determined. Finally, the models were generalized to the whole study area and GC maps were produced. The resulting GC models for both datasets performed in a mediocre way. The best DAP model had a cross-validated RRMSE of 29.8% and the best ALS model had RRMSE of 27.2%. The coefficient of determination (R2) was also better in the ALS model (0.49) than in the DAP model (0.39). The performance of the ALS model was slightly worse than in previous studies using ALS to predict GC. Part of this might be a repercussion of the non-optimal acquisition time of the ALS dataset. For DAP, there were no previous studies. The results of this study suggest that ALS is a more prominent method in mapping forest structure with GC. The DAP proved to be an inexpensive and flexible method of gathering three dimensional information on forests but it had poor canopy penetration abilities which affected the modelling performance negatively.