The 2009 ASHS Annual Conference

For yield and harvest optimization, forecasting of bulking in carrots is critical. Despite being controlled by a plethora of eco-physiological factors, agro-climatological factors have a unique ability to induce and modulate bulking. Modeling the intricacies of this complex relationship demands an approach that offers scalability, flexibility and preciseness. Recently, artificial neural network (ANN) is increasingly used in yield modeling in lieu of multiple linear regressions (MLR) owing to its non-linearity, adaptive learning and diversified algorithm capabilities. Accordingly, this study evaluated the efficiencies of MLR and ANN in modeling the bulking and agro-climatological relationship. Field trials were conducted in 5 geographical locations in Nova Scotia, Canada. Carrot cultivar, Topcut was planted at different seeding dates under 3 seeding rates (40, 55 and 68 seeds/ft). Carrots were harvested at weekly intervals. At each harvest, root girth, height and mass measurements were taken. Of all meteorological and selected agronomic factors, factor analysis deduced growing degree days (base 5 degrees), rainfall (mm) stand count/m, seeding and harvest dates as principal input parameters. Bulking models were constructed using both ANN and MLR techniques. Data from four fields were chosen for model construction and results from the fifth field were used to test the models. The ANN had one input layer, one hidden layer with 6 nodes and one output layer. The model converged after 10,000 epochs and had an r² of 0.89 with internal validation. The MLRs had p values of 0.001 and had a r² of 0.79. Upon testing with independent field data, both ANN and MLR models predicted bulking rates with root mean square error values of 7.35 and 6.90 t/ha. ANN had slightly higher r² (0.75) than that of MLR (0.72).  ANN model showed slightly improved precision and higher flexibility over MLR model.