El Value and limits of GBIF botanical occurrence data in mapping mangrove cover via the KNN classification of Sentinel-2 images

In the current context, which emphasizes the importance of open-access data, this article aims to conduct the first evaluation of the effectiveness of the Global Biodiversity Information Facility’s (GBIF) database for training open-access multispectral satellite image classifiers to map mangrove extent. Six areas in Australia, Benin, Brazil, Colombia, Malaysia, and the USA were selected for their density of mangrove occurrence data and their proximity to other areas with wooded humid vegetation. Three open-access datasets were used: Sentinel-2 images, Google Earth’s high-resolution viewer, and occurrence data downloaded from the GBIF website. First, a k-nearest neighbors (KNN) algorithm was processed and trained with pixels surrounding the GBIF data. Then, the results were compared to those of an unsupervised classification (K-means). The accuracy of the two methods was assessed using error matrices from a double-blind photointerpretation of very high-resolution images obtained from Google Earth. The KNN algorithm achieved kappa indices ranging from 0.85 to 0.94, very similar to those of the unsupervised method (between 0.95 and 0.96). The GBIF possesses a suitable dataset for areas with information, which is why the community is encouraged to fill the main geographical gaps in its database