Modelling the effects of environmental conditions on apparent photosynthesis of Stipa bromoides by machine learning tools

Apparent leaf photosynthesis of the grass Stipa bromoides was measured in the field in two sites of Northern Greece. For predicting apparent photosynthesis from irradiance, temperature and relative air humidity data, we applied and compared two modelling approaches: ordinary statistical modelling and automatic model construction based on machine learning procedures. Ordinary statistical models were constructed based on background knowledge concerning the response of photosynthesis to irradiance. A Michaelis–Menten type light saturation curve was selected among six candidate models and was extended to include air temperature effects. A bell-shaped function of temperature was substituted for the parameter describing the asymptotic maximum photosynthesis. The final model accounted for 67.3% of data variation and was further improved by splitting the data set by experimental site. Site-specific differences were detected regarding the half saturation constant for light and the optimal temperature for photosynthesis. Automatic model construction produced a number of regression trees that enabled a detailed but simple description of the way irradiance, temperature and relative humidity affect photosynthesis. Photosynthesis increases with increasing irradiance, temperature affects photosynthesis when irradiance is close to saturation levels and relative humidity has an effect when both irradiance and temperature are high. There is a threshold value of relative humidity (at about 35%), below which photosynthesis is independent of irradiance within the observed range, decreasing with increasing temperature when temperature is high (>25°C) and increasing with increasing relative humidity when temperature is low (<25°C). The machine learning tools we used provide a very powerful modelling alternative to ordinary curve fitting methods. Their major advantages are the flexibility to select between accuracy and generality and their robustness against outliers and mixtures of differential responses. The models are transparent and easily interpreted. They seem to be able to handle quite complex dependencies among attributes, not requiring prior expert knowledge.