The effect of spatial and radiometric resolutions of aerial images for tree species classification by object-based approach

Abstract

The optimal computational capability for analyzing multispectral aerial images e.g. for fine-scale tree species mapping is often considerably constrained by their enormous data volume. This may be mainly reflected in a reduction in the speed of data processing as well as in their archiving. This research was conducted to explore the effect of alterations in spatial and radiometric resolutions in the quality of object-based tree species classification by UltraCamD aerial images. The study was conducted in three different study sites. Segmentation was firstly implemented on the original images featuring spatial and radiometric resolution of 7 cm and 8-bit, respectively. The optimum segmentation result was then classified. Following this, rescaling in spatial (to 14, 21, 28, 35 and 42 cm pixel size) and radiometric (16-bit to 8-bit) resolutions were conducted, which was followed by classification of the resulted images using the similar segmentation, input bands, features and training and validation data. Based on the conducted accuracy assessment of the resulted classified images, the accuracy was shown to reduce along with a decrease in the radiometric resolution for all of the three areas. However, the trend was shown to be non-uniform when reducing the spatial resolution of the input data. It is concluded that a downscaling of the pixel size down to 4 times coarser than the original pixel size does not notably affect the classification of even-aged or homogeneous forests, while it should be merely conducted with caution in case of natural stands encompassing undisturbed, heterogeneous and diverse groups of species.

Keywords

• UltraCamD
• image segmentation
• object-based classification
• spatial and radiometric resolution