Study on a TCN_BiLSTM_Attention_Based Model for Predicting the Occurrence Levels of Greenhouse pests

Le résumé fourni par la source

With the rapid advancement of agricultural informatization and intelligent technologies, accurate pest risk prediction has become a crucial component of precision pest management. Traditional forecasting models often struggle to capture the nonlinear, high-dimensional, and temporally dependent relationships between environmental variables and pest population dynamics. To address these limitations, this study proposes an integrated greenhouse pests infestation risk-level prediction framework based on a Temporal Convolutional Network–Bidirectional Long Short-Term Memory–Attention (TCN-BiLSTM-Attention) architecture. The approach employs SMOTE and SMOTETomek resampling to balance imbalanced datasets, utilizes the TCN module to extract local temporal features, leverages the BiLSTM network to model bidirectional dependencies, and applies a multi-head attention mechanism to adaptively emphasize key time-step features. Experimental evaluations demonstrate that the proposed model achieves superior performance over benchmark methods (ARIMA, LSTM, GRU, Transformer and TCN), with a precision of 0.87, recall of 0.85, and F1-score of 0.86. Ablation studies confirm that incorporating the BiLSTM and attention modules significantly enhances model accuracy and generalization, while receiver operating characteristic (ROC) analysis indicates an average area under the curve (AUC) exceeding 0.84 across all risk levels. These findings highlight that the TCN-BiLSTM-Attention framework effectively captures complex spatiotemporal dynamics between environmental and pest data, providing a robust and interpretable tool for intelligent monitoring and early warning of greenhouse pest outbreaks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.