KG-CTCN: Knowledge-Guided Causal Temporal Convolutional Networks for Event-Driven Sugarcane Red Rot Forecasting

Abstract: The prevention of Sugarcane Red Rot (Colletotrichum falcatum) outbreaks remains a primary challenge in tropical agriculture, where traditional predictive models often suffer from “monsoon bias”—mistaking seasonal humidity for specific disease triggers. This paper presents KG-CTCN, a Knowledge-Guided Causal Temporal Convolutional Network designed as an event-driven early warning system. Moving beyond daily binary classification, KG-CTCN models 28-day environmental trajectories and integrates agronomic constraints via a cross-modal attention fusion layer. We reconstruct the system's development process, from initial baseline failures compromised by temporal feature leakage to the current deployment-ready architecture. Experimental results on a historical validation set (2019– 2021) demonstrate that KG-CTCN achieves 80% detection of recorded major outbreak events with an average lead time of 12.5 days and a false positive rate of 15.2%—a deliberate design tradeoff reflecting the asymmetric cost structure of outbreak forecasting, where missed detections incur catastrophic crop loss while false advisories impose only marginal spray costs. Robustness tests involving synthetic temporal shifts confirm that the model relies on physical causal signals rather than chronological memorization, marking a significant step toward trustworthy AI in plant pathology.

Keywords: Crop Disease Prediction, Sugarcane Red Rot, Temporal Convolutional Networks, Knowledge-Guided Machine Learning, Early Warning Systems.