Comparison of Mixed Non-linear Models and Support Vector Regression (SVMr) in growth curves of pears

RUBIO, N.; BRAMARDI, S.; GIMENEZ, G.

Barcelona

Conferencia; XXIX International Biometric Conference; 2018

International Biometric Society (IBS)

Horticulture precision is a new requirement to know about the behaviour of the fruits in different climate conditions, variety an size during the development of the fruit, especially in harvest. Nonlinear and non-linear mixed models (NLMM) have been satisfactorily adjusted, based on the logistic model of the third parameterization, which describes the growth of the fruit in diameterfrom the days after full flowering to harvest. Some computational techniques such as Support Vector Machine (SVM) have shown great performance as a classification tool in prediction and presents a version that could be implemented in regression (SVMr), capable of capturing complex relationships in data patterns with lower computational intensity. This work consists in comparing the techniques of NLMM and SVMr, to know if it is possible to use SVMr as an alternative technique in the construction of fruit growth patterns taking advantages of the computational benefits. Data from cultivar Beurre D´Anjou five pears growing cycles were studied: 2005,2006; 2011,2012; 2012,2013; 2013,2014 and 2014,2015, in each cycle 5 trees were selected, taken into account 15 fruits with the following criteria: 5 small fruits, 5 medium fruits and 5 large fruits, each of them were measured in millimeters from flowering to harvest. A mixed non-linear model was fitted from the non-linear equation of the logistic family, considering the random effects of season, plant, fruit size and fruit. At the same time we modelated the correlation of the measurements and the heteroskedasticity found between the seasons. In the case of the SVM, the hyperparameters cost and epsilon were tunning using a grid of possible values for both of them. The calibration was achieved simulating 375 growth curves, simulating the original conditions of the data, considering the estimations obtained in the MNLM. For both techniques, the validation was carried out byadjusting 4 seasons and the 2012-2013 cycle as test. The mean squared error (CME) was applied as a statistical prediction, we calculated the measurement using predicted curve of the fruit size with respect to each fruit. For this research we implemented the statistical language R. The curves obtained by the SVMr had a CME of 7.03 meanwhile the MNLM reached a value of 8.01. In conclusion the SVM could be a useful tool to predict and build the growth curves with the high precision and could be use it in many large databases.