Performance Analysis of Ensemble and Hybrid Deep Learning Models in Predicting Agricultural Chemical Residues and Crop Outcome

Abstract

The contemporary agricultural system has substantially transformed its productivity through extensive deployment of chemical pesticides, herbicides, and fertilizers. However, these practices introduce significant environmental sustainability concerns and human health risks. This paper conducts a comprehensive performance analysis of machine learning (ML) and deep learning (DL) techniques, with emphasis on ensemble and hybrid architectures, applied between 2018 and 2025 to investigate the relationship between agricultural chemical usage, crop yield outcomes, and human health consequences. Through systematic review of 26 peer-reviewed publications selected via PRISMA guidelines, this study evaluates convolutional neural networks (CNNs), long short-term memory (LSTM) networks, hybrid CNN-LSTM architectures, random forests, gradient boosting machines, and probabilistic models. Experimental findings demonstrate that CNN-LSTM hybrid models achieve R² > 0.93 for crop yield prediction, while Graph Neural Networks (GNNs) and Gaussian Process Regression (GPR) attain AUC scores up to 0.91 for health risk assessment. Multi-objective optimization frameworks combining NSGA-II with reinforcement learning yield Pareto-optimal chemical application strategies. Key challenges