Early Prediction and Mapping of Yield in an Orchard Using Color and Shape Features of Apple Fruit.

Early yield prediction is big business in the worldwide fruit industry for both fruit trade and famers alike to plan their picking force, number of bins, storage rooms and sales. Newly developed fruit recognition algorithms are presented for digital images of apple trees at three growth stages of apple fruit, early (July), intermediate (August) and late (September) for 120 cultivar ‘Gala’ apple trees. The great advantage of this approach is that images are acquired in an orchard under natural daylight conditions without extra lighting. A major challenge was the development of the first stage algorithm designed to recognize young, small, light green apple fruitlets against clutter backgrounds including green leaves of commensurate colouration. Back propagation neural network (BPNN) was employed to segment between the spherical shape of the apple fruits and the elongated leaf shape features in order to overcome the ambiguity of light green abaxial leaves and young fruitlets. The task for the second-stage algorithm was to recognize apples undergoing color change from green to red, which was implemented via a combination of color difference and circular shape features. In the third stage, the algorithm employed a color feature of normalized red and green color index (NRG) to recognize bright red apple fruit. Another challenge was the issue of overlapping fruit segmentation relying on the fact that each fruit surface will generate a local bright area by illumination. The coefficient of determination (R2) between algorithm-recognized and visually counted fruit numbers increased from 0.61 for July, to 0.71 for August, and to 0.86 for September. Single tree-based yield was derived from the estimated fruit numbers and diameters. The R2 obtained between the predicted and harvested yields were 0.54, 0.74, and 0.76 for the three consecutive stages, respectively. The yield map of the orchard from the prediction in early July is in good agreement with the yield data collected at harvest. The map of individual tree yield potential as early as July is a great step forward to facilitate more precise, single-tree-based, orchard management.