Hybrid deep ensemble learning for crop disease prediction and climate-resilient disaster management in smart agriculture

Abstract Accurate crop disease prediction and climate-resilient disaster management are essential for improving agricultural productivity and ensuring food security under increasingly dynamic environmental conditions. Existing deep learning approaches typically perform crop disease detection and climate forecasting independently and often rely on deterministic CNN–LSTM architectures without effective multimodal feature fusion, predictive uncertainty estimation, or comprehensive statistical validation. Consequently, their ability to provide reliable decision support under changing climatic conditions remains limited. To address these challenges, this paper proposes a novel hybrid deep ensemble learning framework for integrated crop disease prediction and climate-resilient disaster management in smart agriculture. The proposed framework combines convolutional neural networks (CNNs) for spatial crop image feature extraction with long short-term memory (LSTM) networks for temporal climate forecasting. An attention-based multimodal feature fusion mechanism is employed to integrate heterogeneous crop and climate information, while Bayesian uncertainty estimation and deep ensemble learning are incorporated to improve prediction reliability, robustness, and confidence-aware decision support. Experimental evaluation is performed using multimodal datasets comprising crop images and climate observations collected from the PlantVillage, NOAA, and NASA EOS datasets. The dataset is divided into training, validation, and testing subsets using a 70:15:15 ratio. The proposed framework is evaluated using classification and regression metrics, including Accuracy, Precision, Recall, F1-score, RMSE, MAE, MAPE, and AUC-ROC. To ensure reproducibility, all experiments are repeated five times using different random seeds, and the results are reported with mean $$\pm$$ standard deviation, 95% confidence intervals, and paired Student’s $$t$$ -tests. Experimental results demonstrate that the proposed framework consistently outperforms conventional CNN, LSTM, and CNN–LSTM hybrid models by achieving higher classification accuracy, lower climate forecasting error, improved robustness under varying climate conditions, and statistically significant performance improvements. The proposed framework provides a reliable, scalable, and uncertainty-aware solution for intelligent climate-resilient smart agriculture and early agricultural disaster management.