Neural network modelling of fruit colour and crop variables to predict harvest dates of greenhouse-grown sweet peppers

Abstract

Sweet peppers (Capsicum annuum L.) grown in the greenhouse have irregular yields. Modelling colouration of individual fruit could help growers predict the number of fully coloured peppers that will be ready to harvest within a routine harvest period. We monitored the red, green and blue colour intensities of developing pepper fruit via digital image processing. These colour measurements together with crop phenology and environmental variables were used as inputs into neural network (NN) models to predict days-to-harvest (D-to-H) for ind ividual fruit. When 18 inputs were evaluated, a typical “best” NN model needed only five of the inputs to predict D-to-H (range 0 to 28 d) for red peppers with a R2 of 0.79, a root mean square error (RMSE) of 3.4 d, and an average absolute error (AAE) of 2.5 d. D-to-H were more difficult to predict for yellow peppers, with the “best” model using eight inputs to achieve a R2 of 0.69, a RMSE of 4.4 d, and an AAE of 3.4 d. Light and temperature made little contribution to predictions of D-to-H. NN models with o nly three inputs (Julian day, nodal position of the target fruit and ratio of red:green intensities) could still make useful predictions of harvest maturity. For both red and yellow peppers, the R2 values of NN models were higher than the corresponding R2a (R2 adjusted) values derived from multiple linear regression models. It is concluded that NN have potential to assist greenhouse operators to predict D-to-H of sweet peppers. Key words: Greenhouse production, fruit, colouration, digital imaging, neural networks