Authors: Ying Lu*, University At Buffalo, Le Wang, University at Buffalo
Topics: Remote Sensing, Coastal and Marine
Keywords: large-scale mapping, mangrove forests, Google Earth Engine, training sample collection, one-class classification, region growing, remote sensing
Session Type: Paper
Presentation File: No File Uploaded
Understanding the up-to-date spatial distribution of mangroves at large scales is essential for research on carbon dynamics, climate change, and coastline protection. Supervised classification methods using remote sensing imagery have been demonstrated to be an excellent tool in timely large-scale mangrove forest mapping. For these methods, training sample collection is the most labor intensive process. They are required to be able to describe intraspecific and interspecific variation for land covers. It is a tough task to collect sufficient training samples timely and accurately. Thus, in this study, we developed an automatic training sample collection method to timely map the distribution of mangrove forests at large scales. Based on this new method, we proposed a large-scale mangrove mapping approach with the following three steps. First, using the Mangrove Forests of the World database in 2000 as a baseline, unchanged mangrove areas within 15 years were detected. Second, a region growing method was used to improve the ability of our training samples to describe interspecific variability of mangrove forests. Third, a one-class classification method was used to test the effectiveness of our training samples. Sufficient mangrove training samples with 94.20% accuracy were automatically collected. The overall accuracy of our resultant mangrove forest map is 90.44%. A large part of the approach was implemented in Google Earth Engine, which speeds up data collection and processing. Using automatic training data collection and a one-class classification method, our approach significantly improved the efficiency in large-scale mangrove mapping. It made annual mapping of mangrove forests possible.