Laboratory evaluation of multispectral grape quality grade prediction with compensation for controlled illumination variation
Year: 2026
Type: article
Abstract: Accurate fruit quality assessment is important for selective harvesting in agricultural robotics, where multispectral sensing provides a non-destructive approach but remains sensitive to variability in illumination conditions. This study presents a controlled laboratory evaluation conducted in a dark box under halogen illumination with fixed sensor and sample geometry, in which illumination variation was simulated by adjusting lamp intensity. Prediction models for visually assigned grape quality grades were first established under fixed illumination using an AS7265x multispectral sensor with 18 channels from 410–940 nm, covering the visible and part of the near-infrared range. Spectral data were preprocessed using standard normal variate transformation, standardization, variable importance in projection feature selection, and mixup interpolation. Three machine learning methods, partial least squares regression, support vector regression (SVR), and random forest, were evaluated with nested cross-validation, out-of-fold prediction, and 2000-iteration bootstrap resampling. SVR achieved the best performance, with R 2 = 0.9084 [95% confidence interval: 0.851-0.945], root mean square error = 0.3352, and mean absolute error = 0.2374. To address illumination variability, three compensation methods, direct standardization (DS), extended multiplicative scatter correction (EMSC), and a polynomial per-band fitting method, were implemented and compared across five lamp intensity levels from 20%–100% of rated power. The three methods and the uncompensated condition reached similar mean accuracy across the five levels, with the exception of EMSC, which was lower at every level. DS reduced the spread of the five predictions obtained for the same berry by about 31% relative to no compensation, and gave the highest grade-level accuracy of the four conditions. These results provide laboratory feasibility evidence that DS can improve the consistency of grade prediction across controlled illumination levels, and establish a basis for further validation under natural field illumination toward future robotic harvesting applications.
Source: Engineering Research Express
Authors: Guodong Xiu, Sára Strakošová, Ikuo Mizuuchi
Language: English
Subfield: Analytical Chemistry
Field: Chemistry
Domain: Physical Sciences
SDG: Zero hunger
Open Access status: bronze
OpenAlex ID: W7213509882