Abstract:
Tree height estimation is fundamental in forestry inventory especially in the computation of biomass.
Traditional methods for tree height estimation are not cost effective because of time, manpower and resources involved.
Multiple return LiDAR capabilities offer convenient solutions for height estimations though at equally increased costs. This
study seeks to provide an assessment of the accuracy of Unmanned Aerial System (UAS) stereo imagery in establishing
tree distribution and canopy heights in open forests as an inexpensive alternative. To achieve this, we: generate accurate 3
dimensional surface and bare earth models from UAS data and using these products; establish tree distribution and estimate
canopy heights using data filters; and validate the results using ground methods. A Mavinci Sirius fixed wing Unmanned
Aerial Vehicle (UAV) fitted with a 16 Megapixel camera and flying at an average height of 371 m Above Ground Level
(AGL) was used to image approximately 2 km2 capturing 380 images per flight. An image overlap of up to 85% was
sufficient for stereo generation at a Ground Sample Distance (GSD) of 10 cm for a flight period of 40 minutes. The stereo
imagery captured were processed into orthomosaics and photogrammetric point clouds with an average point density of 23
points per square meters using Structure from Motion (SfM) techniques. Point cloud segmentation revealed tree distribution
patterns in the Ifakara area, with the Near Infrared band proving useful in filtering out trees from non-vegetated areas. From
the tree height estimations and with validation information from 46 sample trees yielded a correlation coefficient, R2=75%.
The study highlights a simplified and cost-effective approach for generation of accurate three dimension (3D) models from
stereo UAS data. With a survey grade GPS/IMU/INS for direct-on-board geo-referencing, limited controls were required
which reduces the cost of the project. With the ease of varying the size of imagery overlap and flying height, imagery with
improved radiometry can be obtained hence improving the determination of tree distribution, and with multi-view image
matching algorithms processing of UAS imagery is made accurate and inexpensive.