Modeling the dynamic changes in Plasmopara viticola sporangia concentration based on LSTM and understanding the impact of relative factor variability.

Reliable disease management can guarantee healthy plant production and relies on the knowledge of pathogen prevalence. Modeling the dynamic changes in spore concentration is available for realizing this purpose. We present a novel model based on a time-series modeling machine learning method, i.e., a long short-term memory (LSTM) network, to analyze oomycete Plasmopara viticola sporangia concentration dynamics using data from a 4-year field experiment trial in North China. Principal component analysis (PCA)-based high-quality input screening and simulation result calibration were performed to ensure model performance, obtaining a high determination coefficient (0.99), a low root mean square error (0.87), and a low mean bias error (0.55), high sensitivity (91.5%), and high specificity (96.5%). The impact of the variability of relative factors on daily P. viticola sporangia concentrations was analyzed, confirming that a low daily mean air temperature restricts pathogen development even during a long period of high humidity in the field.