Mapping tree cover in European cities: Comparison of classification algorithms for an operational production framework

Abstract

In the framework of the Urban Atlas 2012 production, this paper investigated a set of generative models (Maximum likelihood, k-means) and discriminative models (k Nearest Neighbors, Support Vector Machine and Neural Network) to extract urban-tree cover at a European scale. Based on SPOT-5 images and a training on a large coarse resolution dataset, this study tested the performance of these algorithms on 3 cities regarding their geographical location, urban morphology and acquisition dates. Result reveals that discriminative models are more robust than generative ones. It shows that overall accuracy varies from 75% for the k-means classifier to 85% for the neural network. It also shows that neural networks provide the most balanced results (ratio between commission and omission errors) leading to be most suitable algorithm to process different cities with heterogeneous data.