Temporal Vine Water Status Modeling Through Machine Learning Ensemble Technique and Sentinel-2 Multispectral Images Under Semi-Arid Conditions

New challenges will be experienced by the agriculture sector in the near future, especially due to the effects of climate change. For example, rising temperatures could result in increased evapotranspiration demand, causing difficulties in the management of irrigation practices. Generally, an important predictor of plant water status to be taken into account for irrigation monitoring and management is the stem water potential. However, it requires a huge amount of time-consuming fieldwork, particularly when an adequate data amount is necessary to fully investigate the spatial and temporal variability of large areas under monitoring. In this study, the integration of machine learning and satellite remote sensing (Sentinel-2) was investigated to obtain a model able to predict the stem water potential in viticulture using multispectral imagery. Vine water status data were acquired within a Montepulciano vineyard in the south of Italy (Puglia region), under semi-arid conditions; data were acquired over two years during the irrigation seasons. Different machine learning algorithms (lasso, ridge, elastic net, and random forest) were compared using vegetation indices and spectral bands as predictors in two independent analyses. The results show that it is possible to remotely estimate vine water status with random forest from vegetation indices (R2 = 0.72). Integrating machine learning techniques and satellite remote sensing could help farmers and technicians manage and plan irrigation, avoiding or reducing fieldwork.