CATSEM: A Climate-Aware Time-Series Ensemble Model for Enhanced Paddy Yield Prediction

Abstract

Accurate paddy yield prediction remains a vital challenge in agricultural data analytics due to complex climate–soil interactions and regional variability. The proposed Climate-Aware Time-Series Ensemble Model (CATSEM) integrates discrete wavelet decomposition, exponential weighted smoothing, Kalman filtering, and adaptive ensemble learning to capture temporal dependencies in climatic variables. The model preprocesses rainfall, average temperature, and solar radiation through Discrete Wavelet Transform (DWT) for trend extraction, followed by Exponential Weighted Moving Average (EWMA) smoothing and Kalman filtering for signal refinement. Three base learners Long Short-Term Memory (LSTM), XGBoost, and LightGBM are trained on temporally enhanced features, and their outputs are fused using a linear meta-learner. Experimental evaluation demonstrates improved robustness and accuracy with CATSEM. The proposed model offers interpretable temporal insights, emphasizing the dominant role of temperature in yield forecasting. CATSEM serves as a scalable approach for adaptive agricultural planning under climatic variability.