- Coarse Vector Response Regression for long-term monitoring of ecosystems: revealing the causes of regime shifts in a brackish lagoon hal link

Auteur(s): Mante C., Bernard Guillaume, Durbec Jean-Pierre

(Document sans référence bibliographique) 2014-10-00

Ref HAL: hal-01079122_v1
Exporter : BibTex | endNote

Cover is the most frequently used measure for vegetation surveys. Generally, it is coded with the Abundance/ Dominance Braun-Blanquet's integer code, giving rise to data which are generally considered as ordinal. Since each one of these integers is associated with a whole interval of values, we argue that these codes actually convey more information than that of a simple order, and develop this point of view, considering them as imprecise data. To our knowledge, there is no ready-made method to investigate relation-ships between a vector of such responses and several explanatory variables. Consequently, we propose a three-step method for this purpose. These steps are: (1) randomly recover (through a probability associated with the assessor's subjectivity) possible "original numerical responses"; (2) cluster these numerical response vectors according to an appropriate metric, into an appropriate num-ber of groups; (3) average every variable conditionally to the classifier associated with step two, giving rise to "per group regression functions". This method is applied to explain with hydrological variables the abundance variations of Pota-mogeton pectinatus in a brackish lagoon (the Berre lagoon, Provence, France). We reveal the main relationships between P. pectinatus cover and fresh water inputs, salinity and nutrient abundance (nitrate and phosphate); the obtained results are compared to those from Canonical Correspondence Analysis. The proposed method is also tested on artificial data, similar to the original cover data.