Mapping Tree Cover Patterns in an Urban Arboretum from Multispectral Drone Imagery Using Pixel-Based Classification and Object-based Image Analysis

Heidelberg

2026

13 S.

Sonderdruck

13482S

40006045

Remote sensing methods are valuable in mapping small urban forests. Their small size and the species heterogeneity in these forests, however make it difficult to apply traditional aerial and space borne remote sensing data and analysis methods. High resolution imagery from drones can be used for the characterization of small urban forests. But finding efficient and accurate methods of analysing the large volume of data from such imagery remains a challenge. In this study, we compare pixel-based supervised random forest classification and object-based classification of tree cover patterns in a small arboretum forest in south-east Queensland, Australia. Tree cover in this study refers to the total area of forest land covered by trees. Low tree cover areas are defined as locations that are sparsely covered with trees while high tree cover areas are those that are covered by tree canopies above 30%. A multispectral drone imagery with five bands (Blue, Green, Red-Edge, Red, and Near Infra-red -NIR) from September 2023 was used in this study. Additionally, vegetation indices, including Normalized Difference Vegetation Index (NDVI), Normalized Difference Red-Edge Index (NDRE), Green Normalized Difference Vegetation Index (GNDVI), Normalized Difference Red Green index (NDGR) index and Enhanced Vegetation Index (EVI) were computed and used in the image analysis. Specifically, three approaches were adopted in our analyses. Firstly, random forest classification of drone imagery at pixel level and secondly, object-based image segmentation and classification of the imagery. Finally, deep learning method was used to detect tree crowns in the study area. Results show a strong similarity between the two methods of analysis and imply that properly trained machine learning models running at pixel-level could produce accurate results that are comparable to OBIA-generated results of small urban forest characterization. Further, the deep learning method detected 891 contiguous tree crowns in the arboretum forest. The outcomes from our robust applied comparison of these methods, including the deep learning, demonstrate that fine-resolution imagery can be used to characterize small urban forest both at an individual tree crown level and also to show the patterns of tree cover at the forest level.Keywords Object-based image analysis · Deep learning · Tree crowns · Urban forest · Deep forest