Abstract: The implementation of real-time dynamic monitoring of disaster formation and severity is essential for the timely adoption of disaster prevention and mitigation measures, which in turn minimizes disaster-related losses and safeguards agricultural production safety. This study establishes a low-temperature disaster (LTD) monitoring system based on machine learning algorithms, which primarily consists of a module for identifying types of disasters and a module for simulating the evolution of LTDs. This study firstly employed the KNN model combined with a piecewise function to determine the daily dynamic minimum critical temperature for low-temperature stress (LTS) experienced by winter wheat in the Huang-Huai-Hai (HHH) region after regreening, with the fitting model’s R2, RMSE, MAE, NRMSE, and MBE values being 0.95, 0.79, 0.53, 0.13, and 1.716 × 10−11, respectively. This model serves as the foundation for determining the process by which winter wheat is subjected to LTS. Subsequently, using the XGBoost algorithm to analyze the differences between spring frost and cold damage patterns, a model for identifying types of spring LTDs was developed. The validation accuracy of the model reached 86.67%. In the development of the module simulating the evolution of LTDs, the XGBoost algorithm was initially employed to construct the Low-Temperature Disaster Index (LTDI), facilitating the daily identification of LTD occurrences. Subsequently, the Low-Temperature Disaster Process Accumulation Index (LDPI) is utilized to quantify the severity of the disaster. Validation results indicate that 79.81% of the test set samples exhibit a severity level consistent with historical records. An analysis of the environmental stress-mitigation mechanisms of LTDs reveals that cooling induced by cold air passage and ground radiation are the primary stress mechanisms in the formation of LTDs. In contrast, the release of latent heat from water vapor upon cooling and the transfer of sensible heat from soil moisture serve as the principal mitigation mechanisms. In summary, the developed monitoring framework for LTDs, based on environmental patterns of LTD formation, demonstrates strong generalization capabilities in the HHH region, enabling daily dynamic assessments of the evolution and severity of LTDs.
Read full abstract