Estimation of Fruit Load in Australian Mango Orchards Using Machine Vision

The performance of a multi-view machine vision method was documented at an orchard level, relative to packhouse count. High repeatability achieved in night-time imaging, with absolute percentage error 2% or less. Canopy architecture impacted performance, reasonable estimates on hedge, single leader and conventional systems (3.4, 5.0, 8.2 average error, respectively) while fruit load trellised orchards over-estimated (at 25.2 error). Yield estimations were made for multiple via: (i) human count ~5% trees (FARM), (ii) 18 randomly selected within three NDVI stratifications (CAL), (iii) counts (MV-Raw) (iv) corrected occluded using manual CAL (MV-CAL). Across the nine which results all methods available, FARM, CAL, MV-Raw MV-CAL 26, 13, 11 17%, SD 11, 8, 9%, respectively, 2019–2020 season. 10% less 15 20 assessed. Greater estimation occurred 2020–2021 season due time-spread flowering. Use cases tree level data explored context density maps inform early harvesting interpret crop damage, frequency distributions based per tree.