Multi-source temporal fusion and multi-station learning for early warning of diamondback moth infestation in Brassica crops
Year: 2026
Type: article
Abstract: Introduction: crops. Accurate multi-horizon forecasting requires models that preserve short-lived meteorological exposures while integrating thermal timing, spatial dependence, and recent infestation dynamics within a transparent framework. Methods: The Temporal-Geospatial Pest Network (TG-PestNet) forecasts the weekly percentage of inspected plants classified as DBM-infested 1 to 4 weeks ahead at four Guangdong stations, from origin-time information only. It preserves hourly exposure, encodes lookback weeks on a thermal-time axis, couples stations through a wind-informed graph, and protects infestation history in a gated anchor. Evaluation used 344 synchronized weeks, 66 later test origins, 13 history-only comparators, retraining ablations, matched-width controls, and input-matched generic encoders. Results and discussion: TG-PestNet achieved the lowest observed root-mean-square error, mean absolute error, and mean absolute percentage error in all 12 station-metric combinations, with mean reductions of 5.9%-13.0%. Year-ahead forecasts outperformed persistence in all four test years. Removing the history anchor increased error by 14.3% and remained significant after multiplicity correction, demonstrating its central contribution to forecast accuracy. Meteorological information provided complementary predictive value: its occlusion increased the error by 9.3%, and retraining without it increased the macro RMSE by 3.6%. TG-PestNet also substantially outperformed the two input-matched generic encoders, whose errors were approximately one quarter higher. These results establish TG-PestNet as an accurate, reproducible and auditable framework for multi-station pest forecasting, with component-level analyses that identify the dominant contribution of recent infestation history and characterize the predictive roles of high-resolution meteorology and structured spatial information.
Source: Frontiers in Plant Science
Authors: Dong Zhang, Jiale Wang
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $3,801
OpenAlex ID: W7221510543