Accuracy of tree height estimation using digital aerial imagery

Measuring tree height in forest is time and cost-consuming especially in vast areas. Therefore, using digital aerial images are proposed. The main objective of this research is evaluating capability of such images for estimating trees height.

This research was done based on nine UltraCam-D aerial images with 70% end lap overlay and 60% side lap, 7 cm pixel size and scale of 1:8000 in Tehran. The four spectral visible and near infrared bands were fused with the panchromatic band. The images were geometrically corrected using triangulation method and then digital elevation model (DEM) and digital surface model (DSM) were produced using LPS and INPHO software respectively. To estimate the accuracy of DEM, the elevation of 15 points from the 1:1000 topographic map of Tehran were extracted and compared with the DEM elevation values at the same points. After that, calculating 101 trees height, were performed by subtracting values of DSM and DEM for each tree. To check the precision of calculated height, heights of these trees were measured by Sonto in the field and were compared with the estimated height. The trees include various species of Conifers and Deciduous. Comparison of estimated heights (calculated) with ground values (measured heights) was performed using simple linear regression and Coefficient of determination (R2) and root-mean-square error (RMSE). Normalized mean error (NMPE) statistics were also used to evaluate the error rate of DSMs and also the pairwise T-test was used to compare height differences in two groups of broad-leaved and needle-leaved trees.

Due to the RMSE value based on the five control points (0.22 pixels), and the exact correspondence of the corrected images with the topographic map, the reliability of the geometric correction was obtained with acceptable accuracy. The low density of forest cover in the studied area allowed the ground to be seen in the images. This point made it possible to select the right control points and the number of node points. Achieved results, high linear correlation between actual height value and calculated height (R2 = 82%), RMSE (1.35 m) and average difference between calculated and actual height (1.13 m) indicate that estimated height based, has acceptable accuracy. In addition, no significant difference was observed between precision of estimating Conifers and Deciduous height.

Based on the results, this approach could be used operationally to determine height of the trees in relatively flat areas such as urban forest and flat Zagros forest but its capability should be investigated at mountainous areas, high density and high-slope forests.