Time-series analysis of Sentinel-2 satellite images for sunflower yield estimation

Accurate estimates and predictions of sunflower crop yields at the pixel and field level are critically important for farmers, service dealers, and policymakers. Several models based on remote sensing data have been developed in yield assessment, but their robustness-especially in small field scale areas-needs to be examined. Here we aim to develop a robust methodology for estimation/prediction of sunflower yield at pilot field scale using Sentinel-2 remote sensing satellite imagery. We conducted the study in Mezo ˝hegyes, south-eastern Hungary. The Random Forest Regression (RFR), a machine learning technique was used in this research to translate the Sentinel-2 spectral bands to sunflower yield based on crop yield data provided by a combine harvester equipped with a yield-monitoring system. Sentinel-2 images obtained from April to September were used to find the best image for prediction. The satellite image acquired on June 28 was found best and considered further for prediction sunflower yield. A developed training model was tested and validated in 10 different parcels to evaluate the performance of the prediction. We examined the results of the prediction model (predicted) against the actual yield data (observed) collected by a combine harvester. The results demonstrated that using 10 spectral bands from Sentinel-2 imagery the best time to predict sunflower yields was between 85 and 105 d into the growing season during the flowering stage. This model achieved high accuracy with low normalized root means square error (RMSE) ranging from 121.9 to 284.5 kg/ha for different test fields. Our results are promising because they prove the possibility of predicting sunflower grain yield at the pixel or field level, 3-4 months before the harvest, which is crucial for planning food policy.