Hybrid Stacking Ensemble Method For Crop Yield Prediction

Abstract

Food demand is already reaching unmanageable levels with the continuous rise in the world population, which puts added strain on the agricultural sector to increase productivity, not only is there a need to increase agricultural productivity, but it will have to be done in a sustainable way with limited resources and environmental conditions that are subject to changing and in- creasingly unpredictable conditions. Accurate yield prediction is one of the major problems in agriculture today; decisions that impact food security, adequate resource use, economic fo- recasting, and to an extent sustainable practices all rely on accurate yield predictions. In this dissertation, it presents a hybrid approach, H_StackCYP, that combines traditional machine learning and deep learning to predict crop yields. The hybrid model is a stacked ensemble of active learning models that can utilize non-linear predictive models such as Random Forest, XG- Boost, Decision Trees, Convolutional neural networks (CNN), Deep Neural Networks (DNN) and Multilayer perceptron (MLP) models so it is not just linear trends in agricultural data. An important aspect of this work is the anticipation of humidity values as an additional input para- meter which provides more context to the yield estimation phase of the study. The study also investigates complex feasible engineered features and full hyperparameter tuning to optimize each model. The findings from this study demonstrate that the H_StackCYP hybrid model out- performed models applied individually as well as demonstrating decent accuracy, with R2 equal 0.983 in best cases. The project effectively highlights the potential of using ensemble methods for agriculture forecasting and the value of using climate-based features to improve the pre- diction, also it demonstrates the usefulness of the model as a real-world and scalable solution farmers or agriculture planners may leverage in their decision-making to enhance productivity through the use of data-driven methods. Keywords: Crop yield prediction, Hybrid stacking, Ensemble learning, Machine learning, Deep learning, Humidity prediction, Agriculture.