Assessing alternative methods for unsupervised segmentation of urban vegetation in very high-resolution multispectral aerial imagery

To analyze types and patterns of greening trends across a city, this study seeks to identify a method of creating very high-resolution urban vegetation maps that scales over space and time. Vegetation poses unique challenges for image segmentation because it is patchy, has ragged boundaries, and high in-class heterogeneity. Existing and emerging public datasets with the spatial resolution necessary to identify granular urban vegetation lack a depth of affordable and accessible labeled training data, making unsupervised segmentation desirable. This study evaluates three unsupervised methods of segmenting urban vegetation: clustering with k-means using k-means++ seeding; clustering with a Gaussian Mixture Model (GMM); and an unsupervised, backpropagating convolutional neural network (CNN) with simple iterative linear clustering superpixels. When benchmarked against internal validity metrics and hand-coded data, k-means is more accurate than GMM and CNN in segmenting urban vegetation. K-means is not able to differentiate between water and shadows, however, and when this segment is important GMM is best for probabilistically identifying secondary land cover class membership. Though we find the unsupervised CNN shows high degrees of accuracy on built urban landscape features, its accuracy when segmenting vegetation does not justify its complexity. Despite limitations, for segmenting urban vegetation, k-means has the highest performance, is the simplest, and is more efficient than alternatives.