Wheat yield prediction using machine learning, proximal sensing, nitrogen and climate data for Oklahoma and North Africa and the evaluation of NDVI response index consistency using proximal sensors and satellites

Accurate and timely crop yield prediction is critical for optimizing food availability, forecasting economic returns, and setting market prices. It also allows in-season management adjustments, such as optimization of nitrogen (N) fertilization. Machine Learning (ML) have gained prominence for handling big data and modeling non-linear relationships. This study evaluates the effectiveness of using ML with climate variables, canopy sensing data, and N application rates to predict wheat yield in Oklahoma and North Africa. Data were collected from 41 site-years in Oklahoma and 13 site-years across North Africa. The dataset included climate variables, canopy sensing data (NDVI), and pre-plant N application rates. In Oklahoma, the CatBoost model achieved the highest performance, with an R² of 0.81 and an RMSE of 0.63 t/ha. The feature set for Oklahoma was optimized to 13 variables, including NDVI at various growth stages, monthly temperature, rainfall, and evapotranspiration. In North Africa, CatBoost model performed best, with an R² of 0.42 and an RMSE of 0.92 t/ha. The feature set for North Africa was simplified to 14 variables, focusing on NDVI measurements, N rate and climate variables like monthly rainfall and temperatures. This research highlights the potential of integrating ML, remote sensing, weather, and N management for accurate yield prediction. Aside from yield prediction, another important concept for in-season N recommendation is the Response index NDVI (RINDVI), which evaluates crop responsiveness to additional N fertilizer. In a separate study,the response index RINDVI was calculated using NDVI values from GreenSeeker™ and satellite platforms (PlanetScope and Sentinel-2). NDVI data were collected from both the N-rich strip (high N application) and the standard practice strip. High NDVI correlations between GreenSeeker™ and satellites enabled calibration using GreenSeeker™ as a reference. However, even after calibration, GreenSeeker™ consistently displayed a broader range of RINDVI values, leading to differences in N recommendations. PlanetScope showed the closest alignment with GreenSeeker™, with an average underapplication of 4 kg/ha, though deviations of up to 33 kg/ha were observed. The study demonstrated underapplication of N when using satellites, compared to GreenSeeker™ as a reference within the current context of agricultural practices for the N-rich strip.