AutoML and explainable AI (XAI) for rice production systems: Unraveling yield predictors and greenhouse gas emissions in Bangladesh

Explainable artificial intelligence (XAI) offers valuable insights into the factors influencing rice yields and greenhouse gas (GHG) emissions, both of which are crucial for addressing yield gaps and promoting agricultural sustainability. In this study, we employed machine learning models within an automated machine learning (AutoML) framework to predict rice yields and identify the key influencing factors for Boro (the transition from winter to the rainy season), Aman (the rainy season), and Boro (the irrigated winter season) rice in Bangladesh. We utilized model-agnostic interpretation methods to analyze and visualize the impact of these factors on rice yields. The stack ensemble method, which optimally combines multiple machine learning models, outperformed all individual models in predicting rice yields across all seasons. Permutation-based feature importance analysis identified five key factors influencing rice yield: rice type (local, high-yielding variety (HYV), or hybrid), field duration, and the application rates of nitrogen (N), phosphorus (P), and potassium (K) fertilizers. Breakdown plots for additive attributions revealed spatial variations in the average effects of these factors on yield. Moreover, the positive, spatially variable relationships between total GHG emissions and these influencing factors underscored the challenges of mitigating GHG emissions in tropical rice-growing environments.