Prediction of forest soil nutrient status using vegetation

A large number of plots (306) were sampled for floristic data and variables of soil nutrient availability in a wide range of forest conditions in the Vosges mountains (NE France). Half of the data were used to model species presence/absence as a function of nutritional variables using two numerical methods: parametric logistic regression and non-parametric Kernel estimation. Species responses were established for calcium, aluminium, Ca : Al ratio, base saturation, pH and C : N measured in the upper soil horizon. The other half were used to predict these variables with vegetation with three approaches: maximum likelihood, indicator values and ecological groups. More than 80 % of the 122 studied species showed a significant response to one or more nutrition factors. Species often had optima for maxima or minima of the studied values. For each variable, species were more frequent in conditions with good nutrition and/or low toxicity: maximum values were most frequent for Ca, Ca : Al, base saturation, pH and minimum values were most frequent for Al and C : N. For all variables, optima estimated by logistic regression were well correlated with those estimated by kernel estimation: R2 > 0.9. Mean differences between measured values and predictions by vegetation were ca. 2.6 meq for Ca, 1.2 for log Ca, 3.1 meq for Al, 1.8 for log Al, 30 for Ca : Al, 2.1 for log Ca : Al, 0.21 for base saturation, 0.7 for pH and 5.7 for C : N. The quality of prediction as measured by R2 between predicted and measured values varied from 0.60 to 0.14. The quality of prediction decreases from soil base saturation to aluminium according to the next order: base saturation > log Ca : Al > log Ca > Ca > C : N > pH > log Al > Al > Ca : Al. Maximum likelihood and ecological groups methods of prediction were more efficient than the indicator values method. Abbreviations: BS = Base Saturation; EG = ecological group; IV = Indicator value; KE = Kernel estimation; LR = Logistic regression; ML = Maximum Likelihood; RMSE = Root mean square error. Nomenclature: Tutin et al. (1976, 1977).