Prediction of grape yields from time-series vegetation indices using satellite remote sensing and a machine-learning approach

Grape yield prediction is an important tool used by growers to optimize vineyard management and obtain better income. In this regard, the growth stages of grapes play important roles in the evaluation of vineyard production throughout the season. Predicted yield maps allow growers to view spatial variability across fields and to determine the best harvesting time and marketing strategy. Different methods are used to estimate yield; however, large-scale estimation is difficult due to labor and time requirements. Machine learning and satellite remote sensing have the potential to obtain quick and rapid assessments over large areas with lower costs and in shorter timeframes. In this context, the purpose of this research was to develop yield prediction models based on a machine-learning approach using satellite-based time-series images. In this study, Landsat 8 surface reflectance products from 2017, 2018 and 2019 were used to map the satellite-based normalized difference vegetation index (NDVI), leaf area index (LAI) and normalized difference water index (NDWI). Moreover, different growth stages were observed using moving averages and exponential smoothing methods based on the per-pixel values from the satellite imagery. The vegetation indices had particularly close relationships with each other at the time of maximum canopy expansion. To remove the seasonality of the vegetation indices, a moving average was applied to determine one representative mean for each vineyard. The generated models were validated using regression analysis and an artificial neural network (ANN) approach. The results indicated that of all the vegetation indices, NDVI had the highest accuracy (r(2) = 0.79) in 2017 and 2019; however, the LAI accuracy was higher than the accuracies of the other indices (r(2) = 0.79) in 2019. Nevertheless, the artificial neural network-based machine learning results indicated that NDVI had the highest accuracy in 2017 (R = 0.94), 2018 (R = 0.95) and 2019 (R = 0.92) among all the vegetation indices. Thus, machine learning achieved reliable grape yield monitoring across the studied years at local and regional levels. Ground reference yield datasets were used for comparison with the predicted yields. The findings suggest that vegetation indices can be used for calculating site-specific management of vineyards and for predicting yields. The machine-learning methods applied with satellite time-series images can achieve reliable table grape yield prediction models. These integrated models could be used for logistics and decision-making regarding table grape production.