In-season forecasting of within-field grain yield from Sentinel-2 time series data

Abstract

Precise knowledge of cropland productivity is relevant for farmers to enable optimizing managing practices; particularly with the perspective of anticipating crop yield ahead of harvest. The current availability of high spatiotemporal resolution Sentinel-2 satellite data offers a unique opportunity to monitor croplands over time. In this context, the recently introduced kernel NDVI (kNDVI) statistically optimizes the conventional NDVI formulation by applying a nonlinear function to the involved bands, and so maximizes the spectral information extraction. This study proposes a workflow for within-field yield forecasting from Sentinel-2 kNDVI time series analysis focusing on winter cereal croplands in Switzerland over three years, comparing with NDVI as baseline. For a temporally continuous modelling of crop yields, Gaussian Process Regression (GPR) was applied to reconstruct cloud-free time series of the complete crop growing seasons. Following, distinct machine learning regression models (GPR, Kernel Ridge Regression and Random Forest) were developed to forecast yield at any point in time throughout the cropland growing season. The integration of Growing Degree Days (GDD) information as temporal spacing reference of the time series considerably improved the accuracy and consistency of in-season yield forecasting. Training and testing within the same year demonstrated that yield can be accurately forecast approximately 2–2.5 months ahead of harvest, at crops’ anthesis (flowering) phase, with an RMSE up to 0.71 t/ha and a relative RMSE of 7.60%. Although the forecasting accuracy of the models decreased when predicting yield for the unseen years, still satisfactory results were obtained: RMSE = 0.97 t/ha, relative RMSE = 11.47%.

Keywords

• Sentinel-2
• Crop yield forecasting
• Machine learning
• Gaussian process regression (GPR)
• Time series gap-filling
• Growing degree days (GDD)