New tools in forecast of mango crop harvest timing and volume

The management of any supply chain rests on knowledge of stock inventory. For crops, this need involves forecast of harvest timing and volume. This thesis focusses on the use of machine vision to improve forecast of harvest volume and Near Infrared Spectroscopy (NIRS) in context of harvest timing for the Australian mango industry, with attention to issues involved in scaling up machine vision detection of fruit in images to estimation of fruit load of orchards and the issue of calibration robustness for NIRS models. A machine vision system was utilized to estimate fruit load up to six weeks before harvest on fifty four orchards over two seasons. High repeatability was achieved in night-time imaging using a multi-view machine vision method. Yield estimations were made of multiple orchards via: (i) human count of fruit load on ~5% of trees (FARM), (ii) human count of 18 trees randomly selected within three NDVI stratifications (CAL), (iii) multi-view counts (MV-Raw) and (iv) multi-view corrected for occluded fruit using manual counts of CAL trees (MV-CAL). Across the nine orchards for which results for all methods were available, the FARM, CAL, MV-Raw and MV-CAL methods achieved an average percentage error on packhouse counts of 26, 13, 11 and 17 %, with SD of 4, 11, 8 and 11 %, respectively, in the 2019-20 season. Greater error in load estimation occurred in the 2020-21 season due to the time-spread of flowering. However, canopy architecture impacted performance, with reasonable estimates achieved on hedge, single leader, and conventional systems (3.4, 5.0, and 8.2 average percentage error, respectively) while fruit load of trellised orchards was over-estimated. The potential for agronomic interventions to shift fruit load from top of canopy to side of canopy was therefore explored, with the aim of improving fruit visibility from the inter-row and improving harvest efficiency. Upper canopy panicles were removed by pruning in two seasons. In both seasons the proportion of fruit detected by machine vision was significantly increased by the pruning treatment compared to the control (1.51 c.f. 1.19 and 1.56 c.f. 0.88; for 2019/20 and 2020/21 seasons respectively). Fruit were also shifted to the outer canopy as assessed in terms of camera to fruit distances measured using a Time of Flight camera in the 2020/21 season (2238 c.f. 2411 µ distance fruit to camera [mm] for treatment and control row respectively). In another approach, upper canopy vegetative growth was removed at intervals before the floral induction period, preventing formation of upper canopy floral panicles but without decrease in fruit load per tree (control = 61.5; treatments = 59.5 - 74.2 fruit/tree). Short wave near infrared spectroscopy based on Partial Least Squares Regression (PLSR) models has found use in non-invasive assessment of Dry Matter Content (DMC, % fresh weight) of mango fruit. A data set of 4,675 samples acquired across multiple seasons, cultivars and growing regions was assembled for use in training, with harvest populations used as cross-validation groups and data of the fourth season reserved as an independent test set. The fruit physiological stage had the greatest impact on PLSR model performance, compared to cultivar, year or growing region, as estimated using a ‘variable importance metric’. The use of a global ANN model (0.89 %) gave prediction results comparable to specific (cultivar or physiological stage) PLSR models (RMSEP on DMC decreased from 1.01 to 0.88 %). A companion study compared ANN, Gaussian Process Regression (GPR), Local Optimized by Variance Regression (LOVR), Local Partial Least use Squares Regression (LPLS), Local PLS Scores (LPLS-S) and Memory Based Learner (MBL) and two commercially available cloud-based chemometric packages. All of these models gave a better result than a global PLSR model. The best result (lowest RMSEP) was achieved with an ensemble of ANN, GPR and LPLS-S, with the best individual model result achieved by LOVR, with RMSEP of 0.839 % and 0.881 %, respectively, compared to the global PLSR result of 1.014 %. The data set was made public and has been used in multiple publications by other authors, with introduction of deep learning (convolution neural network) models, although direct comparison to the models of this thesis was not possible due to differences in outlier removal. The ANN model was recommended as satisfactory in all categories (model size, prediction speed, model build speed, and prediction statistics) and insensitive to tuning. This ANN model was adopted across handheld spectrometers used in the Australian mango industry in the 2020/21 season.