Tree-level yield mapping for nut trees harvested with shake-catch harvesters, Part II: Uncertainty estimation and tree identification

Abstract

A system that generates yield maps at tree-level resolution for almond orchards and, more generally, shake-catch harvested nut crops is presented in a two-part series of papers. Part I developed an on-harvester yield monitor that uses a laser scanner to profile the continuous nut flow on the conveyor belt and a kernel-based deconvolution was introduced to recover individual-tree yields by fitting a mixture of parametric kernels whose superposition reproduces the measured flow. This paper (Part II) addresses two remaining challenges: (1) quantifying uncertainties in deconvolved, tree-level yields and (2) identifying individual trees that correspond to those yields. Because direct per-tree ground truth after harvest is impractical under mixed-flow conditions, a Monte Carlo method is introduced to experimentally record and weigh single-tree flow profiles and to construct virtual rows in a stochastic harvest simulator. This enabled uncertainty to be analysed systematically over simulated tree-to-tree harvest times and conveyor transport times. A sensor-fusion approach based on dual RTK-GNSS measurements and vibration-sensor timestamps is also introduced to associate harvesting events with individual trees and their deconvolved yield estimates for georeferenced map generation. The results showed that the PMG2-Tukey kernel generally produced lower mean relative errors and lower standard deviations than the other kernel families. Uncertainty was reduced as overlap between consecutive tree flows decreased. The proposed sensor-fusion method enabled the identification of individual trees and the generation of tree-level yield maps.