A statewide urban tree canopy mapping method

Abstract

Maps of urban tree canopy are essential to estimate the magnitude and spatial distribution of ecosystem services, and to determine who benefits from them. Our objective was to develop a repeatable method to map urban tree canopy cover for the state of Wisconsin. We compared two types of imagery (pan-sharpened 1.5 m SPOT-6 and 1 m NAIP aerial photography), three classification algorithms, seven segment sizes ranging from single pixels to 105 m2, three levels of compactness for each segment size, and sixteen subsets of features. NAIP outperformed SPOT. On average across classification algorithms there were no significant differences in map agreement. Pixel-based maps with convolution features performed as well as the best segment-based maps. The best segment size tested was 60 m2 on average, but there was also a local maximum around 15 m2, which suggests a large range of possible segment sizes must be tested to find the best. The model chosen for application to the whole state was a pixel-level Random Forest classifier with 160 features, and the final state-wide map has five classes: tree/woody vegetation, grass/herbaceous vegetation, impervious surfaces/bare soil, water, and non-forested wetland. Overall accuracy for our state-wide map at the pixel-scale was 79.3% (95% CI: 77.5–81%). Errors occurred due to meter-scale heterogeneity in the urban environment which increases the errors due to spatial misregistration and the number of difficult-to-classify mixed and edge pixels. At larger management level scales, mean absolute error (MAE) decreases to about 10% at the Landsat pixel scale and about 6% at the scale of a city block. Our work is comparable to past efforts to map urban tree canopy, and the open source software and use of imagery that is freely available for all cities in the contiguous US make it broadly applicable.