Vine yield estimation from block to regional scale employing remote sensing, weather, and management data

Abstract

Knowledge of the spatial variation in vine yield at different scales is crucial for the wine business, and combined with estimations of vine size variability enables within-block mapping of vegetative-reproductive balance. Remote sensing combined with other data that excludes field sampling appears as an optimal approach for yield estimation for a broad range of scales. In this study, mean yield and factors known to affect yield components were collected for over 8000 blocks, over 18 seasons, in the western oasis of Mendoza, Argentina. Partial Least Squares (PLS) and Random Forest (RF) models were used to analyse the relationship between these factors and yield. The PLS model delivered very poor results, with coefficients of determination lower than 0.08. RF models with 49 to 19 variables produced predictions with coefficients of determination of 0.96 to 0.90, respectively. Some factors traditionally considered important in yield determination, such as trellis, frost occurrence, or planting density had limited influence, whereas location weighed heavily. Results suggest a successful approach to spatial prediction of yield that requires no fieldwork and indicates VRB mapping at block-scale may be possible with these tools. Several improvements to inputs are proposed.

Keywords

• Precision viticulture
• Remote sensing
• Spatial variability
• Machine learning
• Random Forest
• Partial least square regression