| Summary: | Spartina alterniflora (S. alterniflora) after introduced to China, has rapidly expanded along the coastline and become one of the top invasive plants in coastal wetland. While it has been well accepted that phenological information derived from multitemporal remotely sensed data improves vegetation mapping accuracy, previous research primarily relied on scene-based features for invasive plant mapping. In the coastal regions with frequent cloud cover such as South China coast, extracting phenological features at scene level was impossible due to lack of sufficient cloud-free imageries. In this study, we aimed to propose a simple phenological vegetation index (PVI) using pixel-level composition of Sentinel-2 observations with the assist of Google Earth Engine platform. By developing and comparing six PVIs, separability analysis showed that phenological normalized vegetation difference index (PNDVI) of S. alterniflora and other land cover types were more separable than other PVIs and single-season NDVI. Based on the PNDVI, we further proposed supervised and unsupervised Otsu thresholding methods for S. alterniflora extraction. The overall accuracies of supervised Otsu-PNDVI-thresholding method using 10-fold cross validation reached 97.84%, and that of the unsupervised Otsu-PNDVI-thresholding method reached 97.20%. Kappa Z-test statistics showed that both supervised and unsupervised Otsu-PNDVI-thresholding methods yielded statistically similar accuracies as random forest classifiers based on six PVIs, and higher accuracies than scene-based classification. The success of the unsupervised Otsu-PNDVI-thresholding method suggested that the method was practical and operational for S. alterniflora mapping and its expansion monitoring in wide area such as South China coast.
|
|---|
