Abstract
Due to the heterogeneity of urban environments, subpixel urban impervious surface mapping is a challenging task in urban environmental studies. Factors, such as atmospheric correction, climate conditions, seasonal effect, urban settings, substantially affect fractional impervious surface estimation. Their impacts, however, have not been well studied and documented. In this research, we performed direct and comprehensive examinations to explore the impacts of these factors on subpixel estimation when using an effective machine learning technique (Random Forest) and provided solutions to alleviate these influences. Four conclusions can be drawn based on the repeatable experiments in three study areas under different climate conditions (humid continental, tropical monsoon, and Mediterranean climates). First, the performance of subpixel urban impervious surface mapping using top-of-atmosphere (TOA) reflectance imagery is comparable to, and even slightly better than, the surface reflectance imagery provided by U.S. Geological Services in all seasons and in all testing regions. Second, the effect of images with leaf-on/off season varies, and is contingent upon different climate regions. Specifically, humid continental areas may prefer the leaf-on imagery (e.g., summer), while the tropical monsoon and Mediterranean regions seem to favor the fall and winter imagery. Third, the overall estimation performance in the humid continental area is somewhat better than the other regions. Finally, improvements can be achieved by using multi-season imagery, but the increments become less obvious when including more than two seasons. The strategy and results of this research could improve and accommodate regional/national subpixel land cover mapping using Landsat images for large-scale environmental studies.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have