Abstract
Agriculture remains a critical pillar of the Indian economy, yet yield forecasting continues to be affected by climatic uncertainty and diverse environmental conditions. To address these challenges, this study introduces the Robust Lemuria Framework (RLF), a deep ensemble hybrid model that integrates a Deep Belief Network (DBN) for hierarchical and non-linear feature abstraction with a diversified ensemble of Random Forest (RF), J48 Decision Tree (DT), and Naïve Bayes (NB) classifiers for stable prediction consensus. The novelty of RLF lies in its two-stage optimized preprocessing pipeline, which applies DBN-based pre-training to eliminate noise, reconstruct missing values, and refine complex agricultural features through non-linear dimensionality reduction. The framework is trained and evaluated using a decade of multi-regional Indian agricultural data (2010–2020), capturing variations across climate zones, crops, and seasonal patterns. Experimental results show that RLF significantly outperforms existing machine learning and deep learning approaches, achieving 98.99% accuracy, 98.54% sensitivity, 99.35% specificity, and an R(2) score of 0.9994 for yield prediction. These outcomes demonstrate the robustness, scalability, and real-world applicability of the model for agricultural forecasting. Overall, the proposed framework provides a reliable decision-support tool for precision agriculture, contributing to improved crop planning, resource allocation, and policy formulation.