Abstract Purpose Accurate ground-truth data are essential for training reliable remote sensing (RS) models for vineyard yield estimation. However, previous UAV-based studies have mainly relied on single-vine sampling, which can introduce geometric inconsistencies due to canopy overlaps, vine spacing, and pruning configuration in Vertical Shoot Positioning (VSP) systems. This study aimed to assess how the choice of ground-truth sampling method, vine-based versus meter-based, affects the accuracy and robustness of UAV-derived yield models. Methods UAV multispectral imagery was acquired at two phenological stages, BBCH 73 and BBCH 85, over a commercial vineyard trained to a VSP system. Yield data were collected using two ground-truth approaches, a vine-based method, recording the grape fresh weight per vine, and a meter-based method, measuring yield along contiguous one-meter canopy segments. UAV images were processed to derive spectral (NDVI) and geometric (fraction canopy cover, Fc) variables. Three estimation frameworks were evaluated: ( i ) linear regression models; (ii) machine learning algorithms (Efficient Linear, Support Vector Machine, Random Forest, and Gaussian Process Regression); and (iii) a Bayesian inference model. Results Across all frameworks, the meter-based sampling approach outperformed the vine-based one. Linear regressions yielded higher determination coefficients (R²=0.78) and lower nRMSE. ML algorithms achieved R² between 0.70 and 0.76 for meter-based sampling versus 0.15–0.27 for vine-based sampling. Bayesian model trained on meter-based sampled data demonstrated the highest predictive accuracy, achieving an R² of 0.84 with an nRMSE of 20.47% and a mean absolute error of 17.46%. In contrast, the model trained on vine-level data exhibited markedly lower performance (R² = 0.41, nRMSE = 50.08%). At the field scale, the meter-based sampling strategy, when coupled with Bayesian inference, delivered high-fidelity yield estimates with reduced posterior uncertainty. Conclusion Ground-truth sampling design significantly affected UAV-based vineyard yield estimation. Sampling contiguous canopy segments preserves spatial continuity and mitigates geometric bias, enhancing model transferability and reliability across analytical frameworks. This approach provides a robust methodological foundation for scalable RS-based yield modelling in VSP vineyards.
Canicattì et al. (Thu,) studied this question.