A Prediction Model Optimization Critiques through Centroid Clustering by Reducing the Sample Size, Integrating Statistical and Machine Learning Techniques for Wheat Productivity

Machine learning algorithms are rapidly deploying and have made manifold breakthroughs in various fields. The optimization of algorithms got abundant attention of researchers being a core component for deploying the machine learning model (MLM) abled to learn the parameters in significant ways for the given data. Modeling crop productivity through innumerable agronomical constraints has become a crucial task for evolving sustainable agricultural policies. The cross-sectional datasets of 26430 (D1) crop-cut experiments are taken by 2nd-stage area frame sampling, collected from crop reporting service. This research is taken as follows: firstly three more effective numerical optimized datasets are generated (D1, D2, and D3) from D1 by taking the centroid points of features which decrease the sample size; secondly MLM is integrated with the traditional statistical models (TSMs) for multiple linear regression (MLR), and thirdly decision tree regression (DTR) and random forest regression (RFR) are deployed to get the optimized models able to predict the wheat productivity well with 75% datasets to train and 25% to test the model using the evaluation metrics (R(2), RMSE), information criterion (AIC) with weights (AIC(W)), evidence ration (E.R), and decompositions of prediction error. The MLR outperformed for MLM than TSM. The performance capability of MLM and TSM got upswing for generated datasets. RFR got optimized and superperformed for D1, D2, D3, and D4. This study demonstrated strong evidences for deploying MLM for prediction of wheat productivity as an alternative of traditional statistical modeling.