A comprehensive review on wheat yield prediction based on remote sensing

Wheat, one of the most widely cultivated cereals worldwide, is a nutritional source for millions. Accurate mapping of wheat yield is crucial in the planning and decision-making for food security. In recent years, remote sensing (RS) data have been extensively used to predict wheat yields due to their systematic acquisition, broad coverage, and cost-effectiveness. This paper begins by reviewing the different platforms, sensors, and standard features in wheat yield modeling. Additionally, we discuss strategies and methods for feature selection, focusing specifically on monitoring wheat yield using time series data, with MODIS and Landsat datasets found to be the most widely applied satellite data for wheat yield prediction. The review also showed that the normalized difference vegetation index (NDVI), soil-adjusted vegetation index (SAVI), temperature, rainfall, humidity, soil bulk density (SBD), and soil organic carbon (SOC) are among the most widely used vegetation indices (VIs), climate and soil data. In contrast, topography and stress indices are the least applied features for wheat yield prediction. Although different studies have reported various models as the best, review results showed that random forest (RF) and neural networks (NN) are widely used and perform well. Many authors have also identified the flowering stage as the optimal time window for yield prediction using RS data. This review presents a comprehensive comparative analysis of models, highlighting their strengths and weaknesses, along with evaluation criteria based on extensive research, with RMSE, R², and RRMSE being the most commonly used metrics for prediction evaluation. Furthermore, future research directions in this domain are also outlined.