Authors: Ruiliang Pu*, UNIVERSITY OF South Florida, Shawn Landry, University of South Florida, Qiuyan Yu, University of South Florida
Topics: Remote Sensing, Urban Geography, Geographic Information Science and Systems
Keywords: urban forest, Pléiades, object-based image analysis, piecewise masking system, random forest, support vector machine, linear discriminant analysis
Session Type: Paper
Start / End Time: 10:00 AM / 11:40 AM
Room: Maurepas, Sheraton, 3rd Floor
Presentation File: No File Uploaded
In this study, we evaluated the potential of five seasonal high resolution Pléiades satellite images for identifying and mapping urban tree species/groups in Tampa, FL, USA. We assessed the capabilities of individual and combined seasonal images to understand the seasonal effect on tree species mapping accuracy. Seven tree species/groups were mapped, including: sand live oak (Quercus geminata), laurel oak (Q. laurifolia), live oak (Q. virginiana), pine (species group), palm (species group), camphor (Cinnamomum camphora), and magnolia (Magnolia grandiflora). Image-objects (IOs) were used as the species mapping unit. A piecewise protocol separated sunlit and shadow/shaded tree canopy IOs prior to mapping tree species. Accuracy indices of species mapping results were compared among the five individual seasonal images and two dry-wet seasonal image combinations. Shadow IOs were spectrally normalized to sunlit IOs, and the endmember fractions of the species/groups were extracted from the seasonal images using the Mixture Tuned Matching Filtering approach and used as additional features to a set spectral and spatial/textural features. Random Forest (RF), Support Vector Machine and Linear Discriminant Analysis classifiers were tested using selected IO features derived from seasonal Pléiades imagery. Results indicate that late spring seasonal (April) image significantly improved mapping accuracies compared to other seasonal images (p<0.01), and a dry-wet combined seasonal images performed even better, suggesting a significant seasonal effect on tree species identification. The results also demonstrate that RF had the best performance among the three classifiers. Therefore, it is important to choose appropriate seasonal remote sensing data for mapping tree species.