Neural network modelling to predict weekly yields of sweet peppers in a commercial greenhouse

Abstract

The production of greenhouse-grown sweet pepper (Capsicum annuum L.) is irregular with a peak-and-valley pattern of weekly yields. We monitored the yields and environment in a commercial greenhouse in British Columbia over six (2000-2005) growing seasons. Light was defined as cumulative light over the current week, with L_1, L_2, L_3, L_4, L_5 and L_6 representing light over previous weeks. Temperature (AvgT) was defined as the current weekly average of 24-h air temperatures, with T_1, T_2 and T_3 representing temperatures over previous weeks. Inputs were also created for the current weekly yield (Y) and previous weekly yields (Y_1, Y_2, Y_3 and Y_4). Neural network (NN) modelling with up to 21 inputs was used to predict yields 1 wk (Y + 1) and 2 wk (Y + 2) in advance of the actual fruit harvest. Data for five different years were combined for model training with the year to be predicted held separate as a validation set. The best models used 13 inputs to predict Y + 1 with an average R2 of 0.66 over the 6 yr. Y_4, Y-Y_1, Y_1, L_1, Y, Y_3, Y-Y_3 and wk (of the year) were important inputs. The environmental inputs were of lesser importance, which suggests that the cyclic nature of pepper yields is inherent in the pepper biology. Predicting Y + 2 was more difficult with an average R2 of 0.59 over the 6 yr. NN have good potential for predicting pepper yields.