Large-scale spatio-temporal yield estimation via deep learning using satellite and management data fusion in vineyards

Abstract

Crop yield forecasting is essential for informed farm management decisions. However, most yield forecasting models have low spatial resolution, late-season predictions, and lack validation for unseen years or locations. These limitations likely stem from the scarcity of large-scale, high-resolution yield measurement data collected over multiple years, which is uncommon in commercial specialty crop operations. Additionally, these limitations raise concerns about the models’ utility and generalizability under new environmental and management conditions within or across farms. In this study, we develop a spatio-temporal deep learning model to forecast wine grape yield early in the season, utilizing a large dataset with high spatio-temporal resolution (i.e., yield data from ∼5 million grapevines of eight cultivars observed over four years). The model is trained on weekly 10 m RGB-NIR time-series satellite data from Sentinel 2A-B, fused with categorical and continuous management inputs, including cultivar type, trellis type, row spacing, and canopy spacing. We assess the model’s generalizability by examining its performance on data from unseen years and/or locations and at multiple spatial resolutions. Our results show that combining management data with satellite imagery significantly improves model performance on entirely unseen vineyard blocks at 10 m resolution, achieving an R2 of 0.76, a mean absolute error (MAE) of 4.21 tonnes/hectare, and a mean absolute percent error (MAPE) of 13%. We find that cultivars with considerable year-to-year yield variability tend to exhibit lower predictive performance and may benefit from longer time-series observations for model training to encompass a wide range of environmental and management conditions. We also observe improved estimations from the early season in April to the middle of the growing season in June. In conclusion, the yield forecasting and model validation framework established in this study lays the foundation for training spatio-temporally aware deep learning models on exceptionally large yield datasets with high spatio-temporal resolution. We anticipate that such models will become increasingly prevalent as yield monitors are more frequently deployed in specialty crop operations.