A Minimax-Based Intelligent Multimodal Framework for Precision Pesticide Recommendation in Sustainable Agriculture

Abstract

Effective pesticide recommendation plays a crucial role in safeguarding pesticide health while minimizing environmental and economic impacts. Traditional methods of pesticide selection often rely on generalized rules or manual expertise, which may not adequately adapt to dynamic field conditions and diverse pesticide environments. This study proposes a novel, intelligent framework for pesticide recommendation using Integrated Multimodal Data Fusion and advanced learning architectures. The model incorporates time frequency analysis through the Fourier, Discrete Cosine, and U-Transform components, alongside spatial pattern extraction via convolutional layers. These extracted features are enhanced using a Bidirectional Gated Recurrent Unit (BiGRU) and optimized via a hybrid Fuzzy-Augmented Local Dimensionality Reducer (FALDR) that leverages Ant Lion Optimization (ALO) for high-variance feature selection. For pesticide advisory, the refined features are processed through a dedicated Recurrent Frequent Pattern Mining model (RFPMax) that combines the sequential modeling capabilities of Recurrent Neural Networks (RNNs) with transactional insights derived from FPMax algorithms. The model achieves superior precision, recall, and F1 scores across pesticide classes such as Propiconazole, Lambda-cyhalothrin, and Tricyclazole, indicating its robustness and suitability for deployment in precision agriculture. The framework offers a scalable, adaptive, and accurate approach to intelligent pesticide recommendation, contributing to sustainable pesticide protection and optimized farm input usage.