Yield estimation using machine learning from satellite imagery

¹ GMV, Remote Sensing and Geospatial Analytics Division, 28760 Tres Cantos, Madrid, Spain
² Pago de Carraovejas, R&D Department, 47300 Peñafiel, Valladolid, Spain

Abstract

Accurate and early yield estimation (from pea size) allows 1.- Make decisions at field level: green harvesting, irrigation management. 2.- Advance or organise the purchase of grapes from suppliers. 3.- Forecast the volume of wine produced in the campaign that has not yet begun. 4.- Define the quality of the vintage: regular and detailed monitoring of whether, or not, the heterogeneity of the leaf surface, photosynthetic activity or soil moisture observed in the vineyards is as expected at this time, compared with historical values. 5.- Precise control of each vine in production, knowing which vines are no longer productive or should be grubbed up. The Sentinel-2 satellite has generated a time series of images spanning more than six years, which is a great help in analysing the state of permanent crops such as vineyards, where grapes are produced every year. The weekly comparison of what is happening in the current season with what has happened in the previous six seasons is information that is in line with agricultural practices: Winegrowers make the mental exercise of comparing how the vines are developing today with how they developed in previous seasons, with the aim of repeating the years of good yields. In addition, several commercial satellites can now capture images of 50 centimetres pixel resolution or even better, making it possible to check the health of each vine every year. Since 2020, GMV and Pago de Carraovejas have been working together to develop a yield estimation service based on field information and satellite images that feed machine learning algorithms. This paper describes the path followed from the beginning and the steps taken, summarising as follows: 1. - Machine learning algorithm trained with cluster counting and satellite data. 2. - Adjustment of the number of vines in production in each vineyard using very high-resolution imagery. 3. - Machine learning algorithm trained on real production from past campaigns and historical Sentinel-2 time series. The results obtained by comparing the actual grape intake in the winery with the yield estimation range from 91% accuracy in 2020 to 95% accuracy in 2022.