A machine learning workflow for classifying and predicting the annual climatic status of cotton in Xinjiang, China

Abstract

As machine learning applications increase in crop climatic assessment, this study presents an innovative workflow to improve the lack of generality and interpretability in existing research. We developed flexible, rule-based climatic suitability indices (CSIs) using the maximum likelihood method, designed for specific phenological stages, to evaluate dynamic climatic suitability for crops. Based on these indices, we implemented a classification-regression-reclassification strategy to assess the annual climatic status (ACS) and predict short-term climatic suitability. The adaptability of this workflow is evident in its ability to integrate CSIs for various crops, effectively addressing the diverse needs of regional assessments. This study enables the use and comparison of machine learning models, including support vector machines (SVM/SVR), random forest, and gradient boosting trees (XGBoost), to identify the most effective model for classifying and predicting ACS in different regions. Using cotton in Xinjiang as a case study, SVM-XGBoost and SVM-SVR strategies were selected for short-term ACS prediction, achieving accuracy rates between 81.7 % and 91.0 %. From 1991 to 2020, the analysis identified crop yield potential and key factors, with an increase in normal years within suitable areas, indicating climatic adaptability. Sensitivity analysis revealed the influence of temperature suitability on cotton yield, particularly during seedling and boll formation stages in highly suitable areas for cultivation. Moreover, it emphasized precipitation and temperature during sowing to emergence and boll stages in moderately suitable areas. The workflow enhances generality and interpretability, providing a foundation for climate impact research and crop adaptation planning. Future research could incorporate new data sources, additional crop indicators, and various machine learning algorithms to enhance the generality and stability of the model.