Upscaling and downscaling approaches for early season rice yield prediction using Sentinel-2 and machine learning for precision nitrogen fertilisation

• Downscaling and upscaling approaches for rice yield prediction at tillering stage. • Downscaling approach was more accurate in within-field yield variability prediction. • Machine learning models outperformed null benchmark models. • Proposed approach can support rice farmers in precision nitrogen fertilisation. Early season yield prediction could support rice farmers in adopting precision agriculture for nitrogen fertilisation management. Remote sensing and machine learning (ML) can be used to predict and map crop yield during phenological stages relevant to nitrogen application, like tillering in rice, at both within-field and field scales. This study evaluated the transferability of ML models in early season yield prediction through upscaling and downscaling approaches. The effects of two prediction times (tillering and ripening stages) and training/testing set sizes on ML models performance were evaluated over five rice growing seasons (from 2018 to 2022) in northern Italy, using whole-field-average yields and yield maps. Vegetation indices from Sentinel-2 imagery using the Google Earth Engine platform fed five ML algorithms (Cubist-CUB, Gaussian Process Regression-GPR, Neural Network-NNET, Random Forest-RF, and Support Vector Machines-SVM). ML algorithms were trained with yield maps and tested with whole-field-average yields to obtain a downscaling approach, while the opposite was done to obtain an upscaling approach. The downscaling approach showed higher accuracy than upscaling approach. Ripening stage predictions were more accurate than tillering stages, although the downscaling approach showed small differences between tillering and ripening stages. The highest tillering stage accuracy was achieved by SVM for both downscaling and upscaling approaches with 20 % and 27.8 % of Normalized Root Mean Square Error (NRMSE), and 0.99 and 0.99 of Simple Additive Weighting (SAW) score, respectively. Set size and data distribution effected ML models accuracy, with the highest performance achieved by RF and GPR with 0.80 and 1.00 of SAW score for the downscaling and upscaling approaches, respectively. This study demonstrated how ML models and downscaling approach could support rice farmers to calculate the nitrogen dose using the predicted yield at the tillering stage, enabling them to apply a site-specific nitrogen fertilisation based on the within-field yield prediction variability.