Revisiting frameworks for crop yield prediction based on remote sensing: an agricultural perspective

Abstract

Crop yield predictions are essential to anticipate production volumes, reduce price volatility, and optimize logistics, among other purposes. Currently, one of the most widespread methods for building these predictions is based on the use of remote sensing, which provides spectral crop data in a rapid, precise, and non-destructive manner. Since early 1970s, multiple frameworks have emerged to integrate spectral data into crop yield prediction models. In this thesis, we revisited these frameworks and discussed their strengths and limitations. We have structured this thesis as follows. The first chapter focused on presenting the topic and main objectives of the thesis. The second chapter synthesized existing frameworks for crop yield prediction based on remotely-sensed spectral data. We described different groups and subgroups and discussed three key aspects for crop yield prediction frameworks: 1) data requirements for calibration, 2) trade-off between interpretability and predictive power, and 3) spatio-temporal transferability. The groups ranged from models directly based on remotely-sensed spectral data—whether simple or machine learning-based—to those that rely on derived biophysical variables, such as leaf area index, which can be later assimilated into crop simulation models. The third chapter applied this knowledge to a specific case, evaluating the spatio-temporal transferability of wheat (Triticum aestivum L.) yield prediction models at the field-scale, only requiring sowing date and location as user inputs. We used a dataset of 5,265 wheat fields in Argentina (2018–2023) including yield, sowing date, location, and model-estimated phenology. By integrating satellite-derived absorbed radiation with weather variables, we developed Elastic-Net models to predict yield from 50 days before anthesis to 20 days after. The lowest predictive error in non-transfer models occurred at anthesis, approximately 50 days before harvest (0.61 t ha-1; 16%). When transferred, models maintained accuracy in unknown years, but lost accuracy in unknown regions. Additionally, we simulated operational conditions by forecasting yield in future years using models trained only on previous years, obtaining a median error of 0.67 t ha⁻¹ (17%). We discussed model transfer errors in the context of absorbed radiation-to-yield conversion efficiency, which appeared to vary more across regions than years. Finally, in the fourth chapter, we provided the final remarks and offered insights to improve yield prediction models for real-world applications.