Abstract

The invasive pest Hyphantria cunea poses a serious threat to forestry and agricultural ecosystems. To mitigate this threat, accurate and efficient prediction of its distribution is critical for enabling proactive management and targeted control strategies. However, existing traditional statistical models and machine learning approaches exhibit limited predictive performance due to their inability to effectively capture complex spatiotemporal dependencies and process multi-source heterogeneous data. Furthermore, although deep learning has demonstrated progress in pest prediction, current models still suffer from critical limitations including inadequate feature fusion and insufficient critical temporal feature capture. To address this issue, this study proposes a novel Synergistic Fusion Bidirectional Long Short-Term Memory network (SF-BiLSTM). Centered on the synergistic fusion of multi-source heterogeneous data, its architecture consists of three core stages. Firstly, a Bidirectional LSTM (BiLSTM) is utilized to capture complex spatiotemporal dependencies within environmental sequences. Subsequently, dual-path temporal feature refinement is performed: one path employs mean pooling and layer normalization for feature compression, while the other integrates an attention sub-network to focus on critical time steps and conduct weighted fusion. Finally, an adaptive weighted fusion of the dual-path features is achieved through a weight learning network, yielding the final risk prediction. Experiments on the Shandong Province dataset from 2019 to 2024 demonstrate that the SF-BiLSTM model achieves outstanding comprehensive performance, with a Precision of 0.9667 and an F1-score of 0.9496, significantly outperforming traditional machine learning methods such as Gradient Boosting, KNN, and all baseline models. This model successfully reveals the spatiotemporal patterns and risk levels of Hyphantria cunea outbreaks in Shandong Province, providing reliable decision support for precise regional pest prevention and control.