Urban forests are valued as green infrastructure for the variety of benefits they provide across ecological, social, public health, and economic domains. To understand the distribution of trees, maps are required. Many methods of mapping tree cover use expensive and data-intensive datasets such as hyperspectral and LiDAR imagery, making these methods inaccessible to municipal forest managers. Variations in the modelling approaches of existing, freely available methods are not addressed in sufficient enough detail to understand the trade-offs implicit in these variations. Further, characteristics of urban tree canopy mapping (high spatial resolution imagery, heterogeneous urban environments, the importance of the tree class) make it unclear to what extent results derived from other contexts apply. This dissertation addresses uncertainties related to feature selection methods, spatial variability of accuracy assessments, trade-offs involved in the use of different training sampling schemes and sample sizes, sources of classification error, and error in the reference map. Further, this dissertation proposes an optical method for change detection in urban tree canopy maps and compares estimates of tree cover derived from simple image differencing, a correspondence change detection method, and change detection performed with a reference method that incorporates LiDAR data. Random forest-based feature importance was found to be a more effective feature selection method than Kruskal-Wallis test of independence and correlation-based feature selection. The accuracy assessment varies spatially due to low inter-class variability, which can be characterized by unsupervised clustering. Sampling schemes should be determined considering the relative cost of error. Simple random sampling minimized commission error, stratified even sampling minimized omission error, and disproportionate stratified random sampling balanced commission and omission errors. Stratified even sampling was able to produce accurate maps with fewer samples. Maps produced with optical methods had overall agreement between 84.0-87.4% with optical + LiDAR maps when reference errors were corrected. Simple image differencing produced similar estimates of stable tree cover and tree canopy gain, but not tree canopy loss. The correspondence method visually improved parallax-induced errors and yielded high user’s accuracy for the stable tree class at the cost of greater omission error.