Using satellite thermal-based evapotranspiration time series for defining management zones and spatial association to local attributes in a Vineyard

A well-planned irrigation management strategy is crucial for successful wine grape production and is highly dependent on accurate assessments of water stress. Precision irrigation practices may benefit from the quantification of within-field spatial variability and temporal patterns of evapotranspiration (ET). A spatiotemporal modeling framework is proposed to delineate the vineyard into homogeneous areas (i.e., management zones) according to their ET patterns. The dataset for this study relied on ET retrievals from multiple satellite platforms, generating estimates at high spatial (30 m) and temporal (daily) resolutions for a Vitis vinifera Pinot noir vineyard in the Central Valley of California during the growing seasons of 2015-2018. Time-series decomposition was used to deconstruct the time series of each pixel into three components: long-term trend, seasonality, and remainder, which indicates daily fluctuations. For each time-series component, a time-series clustering (TSC) algorithm was applied to partition the time series of all pixels into homogeneous groups and generate TSC maps. The TSC maps were compared for spatial similarities using the V-measure statistic. A random forest (RF) classification algorithm was used for each TSC map against six environmental variables (elevation, slope, northness, lithology, topographic wetness index, and soil type) to check for spatial association between ET-TSC maps and the local characteristics in the vineyard. Finally, the TSC maps were used as independent variables against yield (ton ha^-1) using analysis of variance (ANOVA) to assess whether the TSC maps explained yield variability. The trend and seasonality TSC maps had a moderate spatial association (V = 0.49), while the remainder showed dissimilar spatial patterns to seasonality and trend. The RF model showed high error matrix-based prediction accuracy levels ranging between 86% and 90%. For the trend and seasonality models, the most important predictor was soil type, followed by elevation, while the remainder TSC was strongly linked with northness spatial variability. The yield levels corresponding to the two clusters in all TSC were significantly different. These findings enabled spatial quantification of ET time series at different temporal scales that may benefit within-season decision-making regarding the amounts, timing, intervals, and location of irrigation. The proposed framework may be applicable to other cases in both agricultural systems and environmental modeling.