Techniques for object-based classification of urban tree cover from high-resolution multispectral imagery

A city's trees provide many environmental and social benefits. To ensure the long-term prosperity of its urban forest, a city should have an effective forest management plan that is guided by timely and accurate spatial information. High-spatial-resolution colour-infrared imagery is a commonly available source of forestry information, but accurate automated tree cover extraction in an urban environment remains a challenge. Presented is an effective, semiautomatic, object-based method for urban tree cover extraction, applied to 23 645 ha of 30 cm colour-infrared imagery of London, Ontario, Canada. Detailed methods, including some new techniques, are presented for the empirical selection of segmentation and classification parameters, the selection of subclasses and training samples, rule-based error correction, and image object border smoothing. A majority-voting interpretation of sample points was performed to reduce the subjectivity of the accuracy assessment. A test of the overall classification accuracy using the proposed method on a 2 km x 2 km image tile showed an improvement of 12.8% over that of a traditional maximum likelihood classification. The overall classification accuracy achieved for the entire city was 89.73%, with user's and producer's accuracy for trees of 75.61% and 86.36%, respectively.