PHTFNet-RPM: a probabilistic hybrid network with RPM for tobacco root disease forecasting

Yunhong Bu¹Tingshan Yao²Shaowu Geng¹Renjie HUANG^3*

• ¹Chuxiong Company of Yunnan Provincial Tobacco Corporation, Chuxiong, China
• ²National Engineering Research Center for Citrus Technology, Citrus Research Institute, Southwest University, Chongqing, China
• ³College of Computer and Information Science, Southwest University, Chuxiong, China

Tobacco growers usually face particular challenges in predicting risks of tobacco root diseases due to complex pathogenesis, concealed early symptoms and heterogeneous farm conditions. To solve this problem, we proposed a flexible Probatilistic Hybrid Temporal Fusion Network with Random Period Mask (PHTFNet-RPM) to forecast the disease incidences and disease indices in future multiple days. It allows for a hybrid-structured input with RPM, incorporating configurable static management variables and time-series variables of weather factors and disease metrics, in which RPM simulates the diverse absences of historical observations. Its internal hierarchically aggregated modules learn cross-variable and cross-temporal feature representations to model the complex nonlinear relation between independent variables and target disease metrics. Especially, the probabilistic theory-based uncertainty quantifications of model and prediction results are designed to improve model's credibility and reliability. To verify the proposed PHTFNet-RPM, a large-scale time-series dataset of tobacco root diseases is constructed by organizing 20-year meteorological and disease survey records from Chuxiong Prefecture, Yunnan Province. Extensive comparative experiments demonstrate that our model achieves 4.44–16.43% lower mean absolute error (MAE) than the existing models (LR, SVR, CNN-LSTM, LSTM-Attention etc.), and it can forecast the reliable disease progression trends under different configurations, even only relying on the historical weather observations.