A nearest-neighbours kernel for classification : a case study of in situ two-dimensional plankton images with correction of total volume estimates for copepods ; Un noyau des plus proches voisins pour la classification : application aux images de plancton bidimensionnelles in situ avec correction des estimations de volume total pour les copépodes

Plankton organisms are a key component of the biosphere: they are at the base of marine food webs and are important contributors to biogeochemical cycles, notably of carbon, nitrogen and oxygen. Indeed, phytoplankton captures carbon dioxide from the atmosphere and produces dioxygen; zooplankton contributes to aggregate and export this carbon at depth, where it is sequestered for hundreds of years. This so-called `biological carbon pump' is studied by ecologists to estimate its efficiency nowadays and in the future, in response to climate change. A modern approach consists in studying how the environment is linked with the functioning of ecosystems through `traits' (i.e., individual characteristics) of organisms. For example, a high correlation has been observed between the size distribution of zooplankters and the carbon sequestration efficiency. In situ imaging instruments and large image databases have been built for plankton, allowing taxonomic classification of organisms and quantification of the total volume of each group based on their morphology. The development of automated classification methods has been essential to help ecologists process data. Among them, Artificial Neural Networks (ANNs) have proven to be efficient and accurate, but their decisions are often hard to interpret. On one hand, in this thesis, we put forward the idea that following the transform-then-classify-simply approach of ANNs using a simple, explicit, transform can result in a classifier whose predictions are both interpretable (thus, trustable) and accurate. The proposed transform is defined as a linear combination of optimal, per-class targets, and the classification is performed, like with ANNs, by a nearest-target decision. Furthermore, as a main theoretical result, we establish that the proposed transform defines a kernel associated with the Weigthed-k-Nearest-Neighbor (W-kNN) classifier, and allows interpreting the W-kNN classifier as a member of a larger family of target-based classifiers, which satisfies an optimality ...