Estimating Winter Wheat Yield By Combining Crop Modelling, Data Assimilation and Machine Learning With Sentinel-1 And Sentinel-2 Data For A Sustainable Agriculture Management Tool.

Increasing the sustainability of farming is a key goal in agriculture for high-yielding production areas, like the UK. Often, fertiliser application to a crop is unsustainable and done without prior or contextual information on crop performance. Comparisons of crop performance can provide a key insight for farmers by estimating how well their crops perform against neighbouring fields, highlighting crop seasons where overuse or underuse of fertiliser may have led to anomalous yields. Regional crop yields can also inform higher level decision makers allowing them to take action in regions that consistently underperform compared to a broader setting, and in regions where yields are being negatively affected by agro-climatological changes.

We demonstrate a system that provides field-level end-of-season yield estimates and in-season yield predictions for winter wheat. It uses an ensemble-based data assimilation technique to localise a mechanistic crop model using leaf area index derived from optical Sentinel 2 data. The crop model is driven by weather data, either reanalysis, forecast or a combination, giving our approach sensitivity to climate conditions, while the optical data refines the model at the field level to reflect the observed crop development. We then further refine the crop model parameterisation by using a machine learning regressor with crop phenology metrics derived from Sentinel 1 & 2 observations. The result is an ensemble of crop model realisations localised to a field from which we derive the yield and the uncertainty.

We have validated this approach against a more than 500 UK farmer reported winter wheat yields over 4 years. This method performs well with an RMSE of ~1.2 T ha-1 and an r2 of 0.52. It also performs well over unseen growing seasons, where applying our approach to independent years in a leave-one-out analysis produced RMSEs between 0.97-1.7 T ha-1. This approach can be applied to other crops and in different countries and we have applied this method to both winter wheat and maize in Ghana and China. We have also developed a land use type classification that identifies winter wheat that, combined with our yield estimates, gives insight as to which fields perform anomalously within a region - indicating unsustainable or poor management practices.