Application of artificial neural network for 'Cavendish' banana maturity and chilling injury prediction

R. Saengrayap, S. Chaiwong, J. Rattanakaran
Winter is a critical season for ‘Cavendish’ banana production in northern Thailand. Exposure to low temperatures leads to slow fruit development and occurence of chilling injury (CI) symptoms. Four bunch cover types, i.e., commercial bagging (CC), non-woven (NW) material, waterproof non-woven (WPNW) material and aluminium foil (ALF), were tested in an attempt to prevent CI and assist fruit development. Banana covers were applied from December 22, 2017 to March 17, 2018 (early winter crop) and January 19 to April 5, 2018 (late winter crop). A total of 24 banana trees (6 trees/bunch cover type) were randomly selected in each crop. Eight data loggers were attached inside the bunch cover (2 data loggers/bunch cover type) to monitor temperatures. Another two data loggers were placed on the banana trunk to monitor the surrounding temperature. After harvesting at mature green and ripening stages, fruit quality of the middle and bottom hands of each bunch were determined. CI occurred in all treatments of the early winter crop (average temperature (Tav)=18.34°C, minimum temperature (Tmin)=12.69°C, and maximum temperature (Tmax)=34.71°C), but CI did not occur in the late winter crop (Tav=20.04°C, Tmin=13.02°C, and Tmax=44.71°C). CI resulted in the lower of the red, green and blue (RGB) colour values of the mature green banana peel in the early winter crop comparing to the late winter crop. However, there were no significant differences (P>0.05) in fractal dimension (FD) values since the same CI severity level was observed. An artificial neural network (ANN) model was developed to find relationships among environmental factors, preharvest treatments, and fruit qualities using bagging type, fruit diameter, fruit circumference, RGB values, FD values, ambient and bagging temperature as inputs; while, CI score, fruit roundness, and hand weight were outputs. The 48 data sets were divided for training (60%) and testing (40%). Ten different ANN architectures were tested by varying number of the hidden nodes (2 to 20 nodes). The 14 nodes-one-hidden-layer ANN was the best model, providing the lowest root mean square error values of 0.23, 0.04, and 54 g for CI score, roundness, and hand weight, respectively, with the highest R2 of 0.966. Hence, the ANN shows promise as a tool to understand the complex relationship that could be used to predict fruit quality in banana and other crops.
Saengrayap, R., Chaiwong, S. and Rattanakaran, J. (2019). Application of artificial neural network for 'Cavendish' banana maturity and chilling injury prediction. Acta Hortic. 1245, 29-34
DOI: 10.17660/ActaHortic.2019.1245.4
bunch cover, fruit quality, image analysis, temperature

Acta Horticulturae