Olive-fruit yield estimation by modelling perceptual visual features

Accurate and early olive-fruit yield estimation is a greatly desired objective in oliviculture as a tool to increase profitability and sustainability of the exploitations. This paper presents the design and testing of a novel methodology that, unlike previous proposals using models fed with indirect variables, exploits computer vision techniques to estimate yield from visual features of existing visible fruit, as traditionally done by field experts. The design comprised a neural network fed with 16 descriptors, 8 calculated per canopy face, including the number of visible fruits (1 descriptor), the area of exposed fruit (1 descriptor), and other analytical descriptors aimed at mathematically modelling fruit dispersion (2 descriptors) and aggregation (4 descriptors). The methodology was trained on a set of 37 sample points (74 images), and externally validated on a set of 10 (20 images), manually taken in a super-intensive olive orchard of the Picual Olea europaea L. variety located in Elvas (Portugal), two months prior to fruit harvesting. The total actual yield manually measured corresponding to the external validation set was 173.23kg, providing the methodology an estimated yield of 177.80kg, what implied a root-mean-square-error of 0.9914 kg per sample point, and an overestimation of 2.64%.