Tree height quantification using very high resolution imagery acquired from an unmanned aerial vehicle (UAV) and automatic 3D photo-reconstruction methods

This study provides insight into the assessment of canopy biophysical parameter retrieval using passive sensors and specifically into the quantification of tree height in a discontinuous canopy using a low-cost camera on board an unmanned aerial vehicle (UAV). The UAV was a 2-m wingspan fixed-wing platform with 5.8 kg take-off weight and 63 km/h ground speed. It carried a consumer-grade RGB camera modified for color-infrared detection (CIR) and synchronized with a GPS unit. In this study, the configuration of the electric UAV carrying the camera payload enabled the acquisition of 158 ha in one single flight. The camera system made it possible to acquire very high resolution (VHR) imagery (5 cm pixel−1) to generate ortho-mosaics and digital surface models (DSMs) through automatic 3D reconstruction methods. The UAV followed pre-designed flight plans over each study site to ensure the acquisition of the imagery with large across- and along-track overlaps (i.e. >80%) using a grid of parallel and perpendicular flight lines. The validation method consisted of taking field measurements of the height of a total of 152 trees in two different study areas using a GPS in real-time kinematic (RTK) mode. The results of the validation assessment conducted to estimate tree height from the VHR DSMs yielded R2 = 0.83, an overall root mean square error (RMSE) of 35 cm, and a relative root mean square error (R-RMSE) of 11.5% for trees with heights ranging between 1.16 and 4.38 m. An assessment conducted on the effects of the spatial resolution of the input images acquired by the UAV on the photo-reconstruction method and DSM generation demonstrated stable relationships for pixel resolutions between 5 and 30 cm that rapidly degraded for input images with pixel resolutions lower than 35 cm. RMSE and R-RMSE values obtained as a function of input pixel resolution showed errors in tree quantification below 15% when 30 cm pixel−1 resolution imagery was used to generate the DSMs. The study conducted in two orchards with this UAV system and the photo-reconstruction method highlighted that an inexpensive approach based on consumer-grade cameras on board a hand-launched unmanned aerial platform can provide accuracies comparable to those of the expensive and computationally more complex light detection and ranging (LIDAR) systems currently operated for agricultural and environmental applications.