Machine Learning-based Crop Yield Prediction by Data Augmentation

Abstract

In this study, the effects of dynamic climate and biophysical parameters and static soil parameters obtained from earth observation satellites on cotton yield estimation were examined with four different machine learning algorithms; multilayer perceptrons, long short term memory, quantile regression and extreme gradient boosting (XGBoost). According to the feature space created from climate (temperature, precipitation, etc.), biophysical (leaf area index, vegetation index, etc.) and soil (sand ratio, water permeability, etc.) parameters, the XGBoost approach predicted cotton yield with the highest accuracy. By applying Shapley Additive Global Importance and SHAP to this model, the driving factors of cotton yield prediction were analyzed. As a result of these analyses, the model explains 32% static, that is, soil parameters, and 68% dynamic parameters. The most important dynamic and static parameters were determined as surface soil moisture and clay.