Phytoplankton global mapping from space with a support vector machine algorithm

摘要: In recent years great progress has been made in global mapping of phytoplankton from space. Two main trends have emerged, the recognition functional types (PFT) based on reflectance normalized to chlorophyll-a concentration, and size class (PSC) relationship between cell and chlorophyll-a concentration. However, PFTs PSCs are not decorrelated, one approach can complement other in a task. this paper, we explore several dominant by combining reflectance anomalies, chlorophyll-a concentration other environmental parameters, such as sea surface temperature wind speed. Remote sensing pixels labeled thanks coincident in-situ pigment data GeP&CO, NOMAD and MAREDAT datasets, covering various oceanographic environments. The is with supervised Support Vector Machine classifier trained pixels. This algorithm enables non-linear separation classes the input space especially adapted for small training datasets available here. Moreover, it provides probability estimate, allowing enhance robustness classification results through choice minimum probability threshold. A greedy feature selection associated 10-fold cross-validation procedure applied select most discriminative input features evaluate performance. best classifiers finally daily remote (SeaWIFS, MODISA) resulting PFT maps compared studies. Several conclusions drawn: (1) highlights weight temperature, wind speed variables recognition; (2) show good agreement with phytoplankton distribution knowledge; (3) MODISA seems perform better than SeaWIFS data, (4) probability threshold screens correctly areas smallest confidence interclass regions.