Reliable estimation of the expected yield remains a major challenge in orchards. In a recent work we reported the development of an algorithm for estimating the number of fruits in images of apple trees acquired in natural daylight conditions. In the present work we tested this approach with night-time images of similar apple trees and further adapted this approach to night-time images of mango trees.

Working with the apple images required only minimal re-parameterization of the algorithm and did not require changes of the algorithm itself. Twenty images were used to calibrate the algorithm, and after re-parameterization the number of objects detected by the algorithm corresponded to 79.9% of the number of apples visible in the images. The procedure was tested with approximately 150 images containing close to 7000 apples, and the results were remarkably close to those obtained with the calibration images, namely the number of objects detected by the algorithm corresponded to 80.2% of the number of apples visible in the images. Applying the "correction factor" derived at the calibration stage (1.251=1/0.799) to these images led to an estimate of the number of apples that was within 1% of the number of apples identified by visual inspection of the images.

Mango images were collected in a tropical Australian orchard at ‘stone hardening’ stage at night under artificial lighting.  Images were divided into four sets (20, 80, 10 and 74 images respectively) collected over two evenings.  The fruit in these images did not yet have significant or consistent red coloration (or ‘blush’), and included a high number of split, commercially unviable fruit, making the yield estimation process more difficult.   The data available for the first two sets included on-tree counts at ‘stone hardening’ which is a reliable predictor of actual yield.    Additionally, the fruit in all images were counted by visual inspection.   The analysis of the mango images was more challenging on three accounts: (1) the fruit was elliptical rather than spherical, (2) the fruit color was not uniform and ranged from reddish to green, and (3) there were a large number of very small branches that could cause erroneous segmentation of the fruit. On the other hand, the fruits were larger and the canopy itself was far less dense than in apple trees. The algorithm was modified to handle these issues and calibrated using the first set (20 images). It was then used to analyze the remaining images. The number of objects detected by the algorithm in the calibration set corresponded to 72% of the number of fruit counted by visual inspection. Applying this correction factor (1.39=1/0.72) to the number of objects detected in the remaining images led to an estimate of 5613 fruits, compared to the visual count of 6031 fruits (i.e. estimate was 93% of actual). This result represents a significant step forward compared with previous efforts to analyze the same images using relatively simpler color, texture and shape analysis. In that previous study, the estimate was 84% of actual, and the detections included significant error rates (18%).

Considering the differences between the two datasets in terms of fruits and canopy, the fact that a single algorithm could be used is very encouraging and shows that this approach should indeed make it possible to obtain good yield estimates after calibrating the algorithm locally using only a small number of images (20 in the present study).