Unsupervised Learning and Geostatistics for Vineyard Management Zone Delineation: Integrating PCA, Clustering, and Kriging in Chile’s Maule Valleys

Powdery mildew (Erysiphe necator (Schw.) Burr.) is a pathogen that threatens vineyard sustainability and profitability. This study presents a reproducible framework that integrates unsupervised machine learning with geostatistics to delineate risk and management zones in Vitis vinifera L. vineyards in Chile’s Maule Region. The workflow comprises (i) data preprocessing; (ii) dimensionality reduction via rotated principal component analysis (PCA) to synthesize multisource attributes; (iii) segmentation and dominance assessment from rotated scores to identify key agronomic factors (e.g., vigor and yield); (iv) spatial validation through factor-wise grouping and cross-validation; (v) geostatistical modeling—empirical isotropic and directional variograms, weighted least-squares fitting to extended models, model selection by cross-validation—followed by kriging of principal components and their variances; (vi) clustering-based delineation of management zones projected onto a spatial grid; and (vii) spatial interpolation and fusion to produce discretized backgrounds with contours of the dominant component. Across seasons, retained components explained at least 70% of the total variance. Silhouette coefficients of 0.47–0.56 indicated moderate-to-good separation and stable dominance patterns. Moran’s I was significant in the first two seasons, evidencing spatial dependence and interannual variation. Cross-validated isotropic ranges typically spanned 25–90 m, reaching ~170 m depending on season and component; directional analysis revealed anisotropy with predominant NE–SW (≈45°) continuity. The framework yields continuous severity/incidence maps and coherent management zones, supporting site-specific management and reduced pesticide use in precision viticulture.