Grapevine production worldwide is adversely impacted by climate change, including limited water availability, low-quality or sudden excess of water, and more frequent, severe, and prolonged heatwaves. As a result, grapevine growers require reliable spatial and temporal information on vine water status to adapt practices. This research evaluates the use of Landsat 8 satellite imagery in conjunction with weather data, and a machine learning algorithm (Gradient Boosting Machine) to predict vine water status in large vineyard blocks. The accuracy of predictions was assessed across both space (mapping) and time (forecast) using block-out and date-out cross-validation techniques. The study was conducted over two consecutive growing seasons on a Vitis vinifera, L. cv. Merlot vineyard in Central California. The ground data included measurements of midday stem water potentials, Psi stem and leaf gas exchange (net assimilation, AN and stomatal conductance, gs). Data acquisition was performed in twenty-four experimental units on the same day of the satellite overpasses. The results of the study demonstrate that machine learning is accurate in predicting vine water status spatially within the training measurement dates with low errors (NRMSE Psi stem = 2.7 %, NRMSEgs = 16.2 %, NRMSEAN = 11.2 %) and a high degree of accuracy (R2 greater than 0.8 in the prediction of all three measurements) as assessed by block-out cross-validation. The temporal forecast, assed via date-out cross-validation, proves to be more challenging, although the addition of ground data at one single spatial location improves the date-out performances and allows the NRMSE to reach 6.8 % for Psi stem with R2 of 0.90, 53.4 % for gs with R2 of 0.74, and 25.5 % for AN with R2 of 0.78. The findings from this study have important implications for precision viticulture. They provide an assessment of Landsat 8 imagery, coupled with machine learning, as a means for growers to monitor and forecast vine water status at the field scale. The study highlights the importance of the validation method to ensure the proper use and assessment of machine learning models on agriculture data.
Assessing the spatial-temporal performance of machine learning in predicting grapevine water status from Landsat 8 imagery via block-out and date-out cross-validation
Cianciola V.;Vivaldi G. A.;
2024-01-01
Abstract
Grapevine production worldwide is adversely impacted by climate change, including limited water availability, low-quality or sudden excess of water, and more frequent, severe, and prolonged heatwaves. As a result, grapevine growers require reliable spatial and temporal information on vine water status to adapt practices. This research evaluates the use of Landsat 8 satellite imagery in conjunction with weather data, and a machine learning algorithm (Gradient Boosting Machine) to predict vine water status in large vineyard blocks. The accuracy of predictions was assessed across both space (mapping) and time (forecast) using block-out and date-out cross-validation techniques. The study was conducted over two consecutive growing seasons on a Vitis vinifera, L. cv. Merlot vineyard in Central California. The ground data included measurements of midday stem water potentials, Psi stem and leaf gas exchange (net assimilation, AN and stomatal conductance, gs). Data acquisition was performed in twenty-four experimental units on the same day of the satellite overpasses. The results of the study demonstrate that machine learning is accurate in predicting vine water status spatially within the training measurement dates with low errors (NRMSE Psi stem = 2.7 %, NRMSEgs = 16.2 %, NRMSEAN = 11.2 %) and a high degree of accuracy (R2 greater than 0.8 in the prediction of all three measurements) as assessed by block-out cross-validation. The temporal forecast, assed via date-out cross-validation, proves to be more challenging, although the addition of ground data at one single spatial location improves the date-out performances and allows the NRMSE to reach 6.8 % for Psi stem with R2 of 0.90, 53.4 % for gs with R2 of 0.74, and 25.5 % for AN with R2 of 0.78. The findings from this study have important implications for precision viticulture. They provide an assessment of Landsat 8 imagery, coupled with machine learning, as a means for growers to monitor and forecast vine water status at the field scale. The study highlights the importance of the validation method to ensure the proper use and assessment of machine learning models on agriculture data.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
