Bayesian yield mapping and uncertainty analysis in vineyards using remote sensing data and grape harvester tracking

Yield mapping in viticulture is crucial for optimizing vineyard management. However, it remains constrained by limited yield monitoring sensors on grape harvesters and traditional remote sensing methods, which provide deterministic predictions without explicit uncertainty quantification. This study develops and evaluates a Bayesian hierarchical approach for vineyard yield mapping.MethodsThe model integrates multi-sensor remote sensing and mechanical harvester telemetry data to generate high-resolution, uncertainty-aware yield maps at the farm scale. Over two growing seasons (2022-2023) in a Sicilian vineyard. Normalized difference vegetation index data from unmanned aerial vehicles and Sentinel-2 satellites were combined with telemetry from global navigation satellite systems mounted on a mechanical grape harvester. Multiple model configurations were systematically evaluated using the brms package in R, and sensitivity analyses were performed.ResultsThe best-performing model selected, based on six evaluation metrics (Bayesian R-2, RMSE, MAE, MAPE, and ELPD), achieved an R-2 of 0.83 and RMSE of 0.26 kg/vine on training data, with independent validation showing an R-2 of 0.78 and RMSE of 0.09. The model effectively captured spatial and temporal yield variations. Specifically, it identified a 15% yield decline in 2023 linked to drought and heat stress, while explicitly quantifying prediction uncertainty.ConclusionThis hierarchical Bayesian approach significantly advances vineyard yield mapping, demonstrating robust generalization potential. Integrating additional data from multiple seasons, different vineyards, and grape cultivars could further enhance predictive accuracy, reduce uncertainty in prediction and improve scalability of the framework. Ultimately, this methodology offers valuable insights for precision viticulture, supporting targeted agronomic management and informed operational decisions.