Agrodeepfusion: A Hybrid Deep Ensemble Architecture for Crop and Fertilizer Recommendation

Abstract

The rapid advancements in deep learning (DL) methodologies have brought substantial transformations in agricultural practices, particularly in predicting crop and fertilizer recommendation. Accurate prediction of crop is crucial for optimizing food production and addressing food security challenges. While traditional statistical and machine learning approaches have been widely used for years, their limitations in handling large scale, nonlinear and complex agricultural datasets highlight the need for advanced solutions. Also struggle with generalizability and performance under diverse, complex, and resource constrained scenarios. To overcome these challenges, this study proposes AgroDeepFusion, a hybrid deep ensemble framework that integrates Convolutional Neural Network (CNN), Long Short Term Memory (LSTM), and Deep Neural Network (DNN) models. The proposed architecture employs stacking and bagging ensemble techniques to combine spatial feature extraction, sequential dependency modelling, and nonlinear feature learning. A five year soil dataset collected from Indian agricultural soil laboratories in Tamil Nadu containing 22 crop classes and multiple macro and micro nutrient attributes was used for evaluation, Experimental results demonstrate that AgroDeepFusion achieved 95% prediction accuracy, clearly surpassing standalone deep learning models CNN (86.5%), LSTM (87.5%), and DNN (84.0%) when tested separately. Also outperformed existing Machine Learning and Deep Learning baselines in precision, recall, and F1-score. Comparative analysis shows the ensemble approach’s effectiveness over standalone models. Furthermore, a rule based fertilizer recommendation module enhances the system’s practical usability. This study contributes a novel ensemble architecture and highlights existing gaps in the current literature, offering an effective solution for sustainable agriculture.