Deciduous forest mapping using change detection of multi-temporal canopy height models from aerial images acquired at leaf-on and leaf-off conditions

Discrimination of deciduous trees using spectral information from aerial images has only been partly successfully due to the complexity of the reflectance at different view angles, times of acquisition, phenology of the trees and inter-tree radiance. Therefore, the objective was to evaluate the accuracy of estimating the proportion of deciduous stem volume (P) utilizing change detection between canopy height models (CHMs) generated by digital photogrammetry from leaf-on and leaf-off aerial images instead of using spectral information. The study was conducted at a hemiboreal study area in Sweden. Using aerial images from three seasons, CHMs with a resolution of approximately 0.5 m were generated using semi-global matching. For training plots, metrics describing the change between leaf-on and leaf-off conditions were calculated and used to model the continuous variable P, using the Random Forest approach. Validated at sub-stands, the estimation accuracy of P in terms of root mean square error and bias was found to be 18% and −6%, respectively. The overall classification accuracy, using four equally wide classes, was 83% with a kappa value of 0.68. The validation plots in classes of high proportion of coniferous or deciduous stem volume were well classified, whereas the mixed forest classes showed lower classification accuracies.