Accurate Wheat Yield Prediction Using Machine Learning and Climate-NDVI Data Fusion

Due to exponential population growth, climate change, and an increasing demand for food, there is an unprecedented need for a timely, precise, and dependable assessment of crop yield on a large scale. Wheat, a staple crop worldwide, requires accurate and prompt prediction of its output for global food security. Traditionally, the development of empirical models for crop yield forecasting has relied on climate data, satellite data, or a combination of both. Despite the enhanced performance achieved by integrating satellite and climate data, the contributions from various sources (Climate, Soil, Socioeconomic, and Remote sensing) remain unclear. The lack of well-defined comparisons between the performance of regression-based approaches and different Machine Learning (ML) methods in yield prediction necessitates further investigation. This study addresses the gaps by combining data from multiple sources to forecast wheat yield in the Multan region in the Punjab province of Pakistan. The findings are compared to the benchmark provided by Crop Report Services (CRS) Punjab, with three widely used ML techniques (support vector machine (SVM), Random Forest (RF), and Least Absolute Shrinkage and Selection Operator (LASSO)) by integrating publicly available data within the GEE (Google Earth Engine) platform, including climate, satellite, soil properties, and spatial information data to develop alternative empirical models for yield prediction using data from 2017 to 2022, selecting the best attribute subset related to crop output. The district-level simulated yield data set was analyzed with three ML models (SVM, RF, and LASSO) as a function of seasonal weather, satellite, and soil. The results indicate that combining all datasets using three ML algorithms achieves better yield prediction performance (
$R^{2}$
: 0.74-0.88). Incorporating spatial information and other properties into benchmark models can improve the prediction from 0.08 to 0.12. Random forest outperformed the competitor models with a Root Mean Square Error (RMSE) of 0.05 q/ha and
$R^{2}$
of 0.88. Comparative analysis shows that random forest with 97% and SVM with 93% yielded better results in the study area.