Deep learning based models for grapevine phenology

Context and purpose of the study – the phenological evolution is a crucial aspect of grapevine growth and development. Accurate detection of phenological stages can improve vineyard management, leading to better crop yield and quality traits. However, traditional methods of phenological tracking such as on-site observations are time-consuming and labour-intensive. This work proposes a scalable data-driven method to automatically detect key phenological stages of grapevines using satellite data. Our approach applies to vast areas because it solely relies on open and satellite data having global coverage without requiring any in-field data from weather stations or other sensors making the approach extensible to other areas.

Material and methods - we leveraged historical phenological observations and developed a supervised deep-learning model that uses the land surface temperature estimated by the Copernicus Sentinel-3 satellite to estimate the current phenological stage at the parcel level. We compared the performances of our model with traditional approach based on Growing Degrees Days (GDD).

Results – we train our algorithm on manually collected phenological observations of four winegrape cultivars in three European vineyards (Italy, Spain, and Portugal) from 2017 to 2022. Preliminary results indicated that our deep learning phenology model outperforms the traditional methods based on GDD, decreasing the Mean Absolute Error from 33.8 to 7.8 days (-76.5%).