Calibration of a growth model for tomato seedlings (TOMSEED) based on heuristic optimisation

• Development of a mechanistic growth model for tomato seedlings.
• Model calibration used genetic algorithms, simulated annealing and tabu search.
• Genetic algorithms were the most successful heuristic approach for model calibration.
• Model validation on real measurements showed good results.
• Number of leaves and shoot length and thickness were modelled more accurately.

A mechanistic growth model for tomato seedlings cultivated in unheated beds is developed, based on modifications of existing tomato growth models. Photosynthetically active radiation at crop level, air temperature and CO2 concentration are taken into account, while simulated variables include dry weights of leaves, shoot and root, leaf area index (LAI) of the seedlings, number of leaves per plant and finally, shoot length and thickness. Model calibration is formed into an optimisation problem, taking into account model errors of the first five simulated variables, i.e., dry weight of leaves, shoot and root, LAI and number of leaves per plant. Three heuristic optimisation algorithms are explored during model calibration: genetic algorithms, simulated annealing and tabu search. Genetic algorithms proved to be the most successful approach, resulting in an overall average deviation between simulated and measured values of around 16%. The calibrated model is tested and validated on measurements not used for calibration, showing a satisfactory performance in modelling most seedling characteristics, like the number of leaves per plant, the shoot length and thickness and the dry weight distribution, while it is not so accurate in predicting of other features like leaf and shoot dry weight and LAI. The tomato seedling characteristics that are mainly related to seedling quality (shoot length and dry weight, and number of leaves) were satisfactorily modelled with an average deviation between measured and simulated values around 16% for most of the simulation period. Finally, possible improvement strategies for future research are also discussed.