Miami-Dade County Urban Tree Cover 2014

The tree cover map for the urban development area of Miami-Dade County was produced from 2-m resolution multi-spectral World View 2 data. The tree-cover map includes eight additional land-cover classes: Grass, Bare Ground, Wetland, Water, Buildings, Impervious Areas, Railroad and Cropland. Method Steps Data Source: WorldView-2 (eight band spectral resolution, 2m spatial resolution) data acquired on 2011-05-09, 2012-12-27, and 2014-02-16. Pre-Processing: All images were geometrically corrected using the rigorous orthorectification module in ENVI. All images were atmospherically correct with the Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) procedure in ENVI. Land-Cover Detection: Multi-spectral reflectance values were used in the classification of nine land cover classes: Buildings, Concrete, Asphalt, Truck Trailers on Asphalt, Bare Ground, Grass, Tree & Shrub, Wetland and Water Training samples for each of the nine classes were digitized from aerial photography and interpretation of WV spectral signatures. The initial land-cover detection was performed using the random forest classification algorithm (Liaw & Wiener, 2002; Svetnik et al., 2003) in the caret R-package (Kuhn & Team, 2014), on the basis of spectral signatures extracted for each training sample form the WV data. Reclassification of Land-Cover: Using vector data layers, provided by Miami-Dade County a post-detection expert classifier evaluated each pixel to belong to either of the final nine classes. Vector layers that were used for the reclassification are large buildings (polygons), small buildings (points buffered with a 5m radius), edge of pavement (polylines converted to polygons), railroads (polylines buffered at 3 m distance), water bodies (polygons), and agricultural areas (polygons). Expert-Classifier Rules: Asphalt, asphalt with trailers, concrete and areas covered by the edge of pavement were reclassified as impervious except for shrubs and trees. Small buildings dominated over all classes except for shrubs and trees, tall buildings dominated all other classes. Cropland dominated over grass and bare ground, but not over shrubs and trees. Morphological Filter: The final map, in order to remove spurious pixels and areas smaller than the class-specific minimum mapping unit (MMU), was spatially filtered using a 4-edge kernel applying a nearest neighbor replacement method. The final filtered classes and their minimum mapping units in square meters: Impervious (40), Roads / Railroad (40), Bare Ground (20), Buildings (8), Grass (20), Trees / Shrubs (8), Wetland (200), Water (200), Cropland (200). Map Accuracy: A design-based accuracy assessment of land-cover class stratified random samples (N = 531; multinomial distribution sampling based on a 95% confidence) estimated the bias adjusted overall accuracy of the map to be 90.1%, with a standard error of 1.7% which means that the 95% upper and lower confidence of the true accuracy is estimated to be between 86.7% and 93.5%. Class-specific map accuracies ranged from 88.1 ± 4.2% for tree canopy to 100% for the street and railroad class. Buildings were mapped with an adjusted accuracy of 93.2 ± 3.3%, grass at 88.2 ± 4.2%, bare ground at 89.4 ± 4.0%, and impervious at 84.8 ± 4.7%. In Table 4, grass was predominantly misclassified as trees and shrub and vice versa at 8.5%, bare ground as impervious (5.1%), and impervious as buildings (8.5%).