A Distribution Graph Guided Network with Dual Track Self-Supervised Strategy for Tobacco Pest Time Series Forecasting

作者：Liming Zhu, Xu Kong, Ming Li, Ting Qin, Yan-bo Chen, Junyu Chang, Xu Chen · 年份：2024 · DOI：10.1145/3704558.3707079 · 研究领域：Advanced Chemical Sensor Technologies、Air Quality Monitoring and Forecasting、Spectroscopy and Chemometric Analyses

Tobacco pest is one of the main factors that harm tobacco quality in tobacco factories, leading to economic losses. Accurate prediction of pest distribution is crucial for the management and production of tobacco factories. Nevertheless, conventional time series forecasting techniques prove inadequate in capturing sufficient information for fine-grained forecasting tasks involving district-variable-level and day-sampling-level tobacco pest datasets. This limitation arises from the lack of consideration for the knowledge pertaining to the migration and reproduction patterns of tobacco pests. In this work, a dual track self-supervised and distribution graph guided network (DTSSDGN) is proposed to handle this problem. First, a distribution graph guided feature extraction module is designed to help capture the internal and external patterns of pest distribution. Subsequently, a two-stage training strategy is devised, comprising a dual-track self-supervised training phase to acquire features that possess a comprehensive understanding of the pest distribution trend, followed by a fine-grained fine-tuning phase that leverages this knowledge to achieve accurate forecasting outcomes. The effectiveness and superiority of the proposed method are illustrated on a real tobacco pest distribution dataset.