@incollection{fdi:010063840,
  title = {{P}hytoplankton global mapping from space with a support vector machine algorithm},
  author = {{D}e {B}oissieu, {F}. and {M}enkes, {C}hristophe and {D}upouy, {C}{\'e}cile and {R}odier, {M}artine and {B}onnet, {S}ophie and {M}angeas, {M}organ and {F}rouin, {R}.},
  editor = {},
  language = {{ENG}},
  abstract = {{I}n recent years great progress has been made in global mapping of phytoplankton from space. {T}wo main trends have emerged, the recognition of phytoplankton functional types ({PFT}) based on reflectance normalized to chlorophyll-a concentration, and the recognition of phytoplankton size class ({PSC}) based on the relationship between cell size and chlorophyll-a concentration. {H}owever, {PFT}s and {PSC}s are not decorrelated, and one approach can complement the other in a recognition task. {I}n this paper, we explore the recognition of several dominant {PFT}s by combining reflectance anomalies, chlorophyll-a concentration and other environmental parameters, such as sea surface temperature and wind speed. {R}emote sensing pixels are labeled thanks to coincident in-situ pigment data from {G}e{P}&{CO}, {NOMAD} and {MAREDAT} datasets, covering various oceanographic environments. {T}he recognition is made with a supervised {S}upport {V}ector {M}achine classifier trained on the labeled pixels. {T}his algorithm enables a non-linear separation of the classes in the input space and is especially adapted for small training datasets as available here. {M}oreover, it provides a class probability estimate, allowing one to enhance the robustness of the classification results through the choice of a minimum probability threshold. {A} greedy feature selection associated to a 10-fold cross-validation procedure is applied to select the most discriminative input features and evaluate the classification performance. {T}he best classifiers are finally applied on daily remote sensing datasets ({S}ea{WIFS}, {MODISA}) and the resulting dominant {PFT} maps are compared with other studies. {S}everal conclusions are drawn: (1) the feature selection highlights the weight of temperature, chlorophyll-a and wind speed variables in phytoplankton recognition; (2) the classifiers show good results and dominant {PFT} maps in agreement with phytoplankton distribution knowledge; (3) classification on {MODISA} data seems to perform better than on {S}ea{WIFS} data, (4) the probability threshold screens correctly the areas of smallest confidence such as the interclass regions.},
  keywords = {{TELEDETECTION} {SPATIALE} ; {DONNEES} {SATELLITE} ; {PHYTOPLANCTON} ; {CLASSIFICATION} ; {ALGORITHME} ; {COULEUR} {DE} {L}'{OCEAN}},
  booktitle = {{O}cean remote sensing and monitoring from space},
  numero = {9261},
  pages = {92611{R} [14  en ligne]},
  address = {{B}ellingham {WA}},
  publisher = {{SPIE}},
  series = {{P}roceedings of {SPIE}},
  year = {2014},
  DOI = {10.1117/12.2083730},
  ISSN = {0277-786{X}},
  URL = {https://www.documentation.ird.fr/hor/fdi:010063840},
}