Canopy-based Classification of Urban Vegetation from Very High-Resolution Satellite Imagery

Urban trees are essential as they provide services in terms of air pollution mitigation, freshness, biodiversity, and citizens’ well-being. Accurate data on location, species, and structural characteristics are essential for quantifying tree benefits. For a realistic and proper quantification of the benefits of urban vegetation in terms of providing ecosystem services at city scale, a consistent inventory of vegetation within residential and public areas, is needed. However, the cost of measuring thousands of individual trees through field campaigns can be prohibitive and reliable information on domestic gardens is lacking due to difficulties in acquiring systematic data.

 

The main objective of this study was to investigate the suitability of very-high resolution satellite imagery for detecting, delineating, and classifying the dominant plant species in both public and private urban areas. The detection of individual trees and species differentiation are challenging in cities, as trees can be isolated, lined up or grouped in patch, with a wide range of plant species, high spectral similarity of vegetation types, and high-density stands, trees in the shade, trees with low spectral contrast, and due to the complexity of the urban environment (buildings, shadows, open courtyards). To overcome these constraints, a canopy-based classification was developed with the selection of new relevant spectral and texture-based features for each tree species and herbaceous areas.

 

A pan-sharpening approach and stepwise masking protocol from WV-2 imagery were used to separate vegetated and non-vegetated areas, tree, and non-tree canopy, over the study areas prior to tree species mapping. The shadows of the trees, but also the shadows of the objects (e.g., buildings) were correctly removed within residential yards. Then, we performed a multispectral procedure of object-based classification using Random Forest classifier with different textural features extracted from tree canopy and grassland (lawn/turf) to identify and map dominant types of vegetation. Four spectral bands (blue, green, yellow, red) and four texture features (i.e., energy, entropy, inverse difference moment, Haralick correlation) were found to be the most efficient attributes for canopy-based classification from WV-2 images.

In both study areas, about 420,000 and 555,000 canopies were successfully classified in Aix-en-Provence and Florence with about 85% in private lands and not under municipalities supervision. We also detected 1,157 and 5,438 herbaceous areas in Florence and Aix-en-Provence, respectively. The number of canopies not classified is very low, i.e., 66 out of 419,399 tree canopies were not classified in Aix-en-Provence (< 0.02%) and 4,030 out of 554,603 tree canopies in Florence (< 0.7%). In the two study areas, Aix-en-Provence (France, 50km²) and Florence (Italy, 80km²), 22 and 20 dominant species were identified and classified with an overall accuracy of 84% and 83%, respectively. The highest classification accuracy was obtained for Pinus spp. and Platanus acerifolia in Aix-en-Provence, and for Celtis australis and Cupressus sempervirens in Florence.