Modelling hydrocarbon contamination in agricultural and environmental systems using hybrid CNN-BiLSTM deep learning: A review

Abstract

Hydrocarbon contamination of agricultural soils, surface water, and groundwater represents one of the most severe environmental threats to food security and ecosystem integrity, particularly in oil-producing regions. Accurate and timely modelling of contaminant distribution, migration dynamics, and concentration levels is critical for risk assessment, remediation planning, and regulatory compliance. This review critically examines the application of hybrid Convolutional Neural Network - Bidirectional Long Short-Term Memory (CNN-BiLSTM) architectures for modelling hydrocarbon contamination across agricultural and environmental contexts. A systematic literature search covering the period 2015-2024 identified 47 relevant studies from peer-reviewed journals, conference proceedings, and technical reports. The review evaluates CNN-BiLSTM frameworks with respect to data input types (in-situ sensor arrays, remote sensing imagery, hyperspectral data, geochemical profiles), architectural configurations, training strategies, and performance benchmarks. Results demonstrate that CNN-BiLSTM consistently outperforms standalone CNN, LSTM, support vector machine, and random forest models, achieving average improvements of 12-28% in prediction accuracy (R² > 0.93) and 18-35% reduction in root mean square error across benchmark datasets. Key advantages include the architecture's capacity for simultaneous spatial feature extraction and bidirectional temporal sequence learning, enabling robust prediction of contaminant plume evolution, soil total petroleum hydrocarbon (TPH) concentrations, and benzene-toluene-ethylbenzene-xylene (BTEX) groundwater levels. Challenges related to data scarcity, transferability across soil types, interpretability, and computational cost are discussed. The review concludes with a research agenda identifying priority gaps and emerging directions, including physics-informed neural networks, federated learning for multi-site contamination monitoring, and integration with precision agricultural management systems.