Improving wheat yield prediction integrating proximal sensing and weather data with machine learning

Abstract

• Ensemble learning achieved the best performance for field-scale wheat yield prediction. • Normalized Difference Red Edge Index was the most contributory features. • Average temperature, minimum temperature, and relative humidity were key weather data. • The low temperature and dry air in winter and spring are limiting factors in wheat yield. Accurate and timely wheat yield prediction is of great importance to global food security. Early prediction of wheat yield at a field scale is essential for site-specific precision management. This study aimed to develop an in-season wheat yield prediction model at field-scale by integrating proximal sensing and weather data. Nine multi-N rates field experiments were conducted at five sites involving different wheat cultivars from 2010 to 2020. Proximal sensing data were collected from a Crop Circle sensor at the stem elongation stage and weather data were collected from 30 days before planting to the flowering date. Eleven statistical and machine learning (ML) regression algorithms were adopted, along with two aggregation intervals (disaggregated or aggregated data) and two feature selection methods (based on Pearson Correlation Coefficient or Recursive Feature Elimination). The results revealed that the ensemble learning models (Random Forest, eXtreme Gradient Boosting) achieved the best overall performance (R 2 = 0.74 ∼ 0.78, RMSE = 0.78 ∼ 0.85 t ha −1 ). Feature importance analysis showed that Normalized Difference Red Edge Index (NDRE), average temperature, minimum temperature, and relative humidity were the most contributory features, especially from the planting date to the stem elongation date (for weather features). The aggregation approach and feature selection method did not significantly affect the yield prediction performance for the seven ML methods. This study introduced a promising framework that complements county-scale models and provided insights into understanding yield responses to environmental conditions. The best prediction model can be applied for guiding real-time sensor-based precision fertilization.