A Knowledge-Embedded Machine Learning Approach for Predicting the Moisture Content of Forest Dead Fine Fuel

The dead fine fuel moisture content (DFFMC) directly affects forest fire occurrence and spread. Accurate DFFMC prediction is key to estimating forest fire risk and behavior. The well-fitting machine learning (ML)-based meteorological factor regression models are a focus of DFFMC prediction modeling. Nevertheless, this method’s reliance on a considerable amount of training data and limited extrapolation hinders its potential for extensive implementation in practice. To improve the prediction accuracy of the model in the context of limited training data volumes and interspecies and spatial extrapolated predictions, this study proposed a novel DFFMC prediction method based on a knowledge-embedded neural network (KENN). By integrating the partial differential equation (PDE) of the meteorological response of forest fuel moisture content into a multilayer perceptron (MLP), the KENN utilizes prior physical knowledge and posterior observational data to determine the relationship between meteorology and moisture content. Data from Mongolian oak, white birch, and larch were collected to evaluate model performance. Compared with three representative ML algorithms for DFFMC prediction—random forest (RF), long short-term memory networks (LSTM), and MLP—the KENN can efficiently reduce training data volume requirements and improve extrapolation prediction accuracy within the investigated fire season, thereby enhancing the usability of ML-based DFFMC prediction methods.