Using pixel- or superpixel-based segmentation for efficient classification of green roof digital images and rapid estimation of plant species abundance
Résumé
Green roofs are “nature-based solutions” that may provide numerous ecosystem
services in the context of urban greening. Vegetation development in strong
interactions with green roof media and microbial communities play a key role in the
ecosystem multifunctionality of these green infrastructures. Consequently, research
on plant selection has increased in recent years in order to expand the list of species
able to grow on green roofs in cold or arid climates, with special efforts to clarify the
relationships between species diversity, functional traits and ecosystem services.
There is still a need, however, for carrying out long-term vegetation survey to avoid
premature conclusions regarding plant adaptation as well as to better understand the
ecosystem functioning and ecological trajectories of these engineered ecosystems.
Nevertheless, vegetation cover determination on green roofs may rapidly become
time-consuming and quite tedious, especially in the case of densely covered mixed
species roofs or due to plant species (sometimes overlapping) with similar leaf shape
and colours (e.g. sedums). Therefore, the aim of this work was to investigate the
ability of two freely available plugins, which were recently developed for the opensource
image analysis software Fiji (a distribution of the free ImageJ platform, initially
dedicated to biological image analysis), to rapidly and efficiently perform supervised
machine-learning for the classification of green roof vegetation from photographs.
Two workflows are thus described, depending on the use of the “Trainable Weka
Pixel Segmentation” or the “Trainable Superpixel Segmentation”, and compared with
the traditional “point frame measurement” method for plant abundance determination
at the species level. Both machine-learning based methods performed very well
regarding image classification efficiency and reproducibility (using fast random forest
algorithm), with the best results obtained with a minimum image label number of 16
per object category to be classified (pouzzolane, moss, sedum and grass species).
Moreover, preliminary image segmentation into superpixels with jSLIC (Simple
Linear Iterative Clustering) with an initial grid size of 1 (and a minimum label number
of 4 per object category to be classified) allowed for highly satisfactory classification
results in a few seconds only, which would facilitate the use of such imagery tool for
routine survey.