基于时空深度学习的土壤温度预测可解释性研究

土壤温度是重要的陆面变量。其时空变化受多种环境因素影响，这给土壤温度的准确预测带来巨大挑战。以机器学习为核心的数据驱动方法，是土壤温度预测研究中的重要手段，也是对传统物理过程模型的重要补充。目前基于机器学习的预测模型和其可解释性仍存在诸多不足。为此，本项目提出一种可解释的时空深度学习模型，在加强土壤温度时空特征学习能力时，融入时空可解释组件，分析土壤温度预测模型的运作机制，进而为地面自动气象监测系统提供多种决策依据。本项目有望在基于机器学习的土壤温度预测模型、土壤温度的陆面过程模式参照等方面取得开拓性成果，对提高土壤温度预测能力，阐明土壤温度时空变化的物理机制有重要的科学价值。

Soil temperature is an important variable in the fields of Earth sciences. The temporal and spatial variation of soil temperature is affected by many factors, and it poses immense challenges for the soil temperature prediction. For soil temperature prediction, data-driven machine learning method is an important research, and it can be an important complement to physics-based process models. However, there are many shortcomings in predictive models and their interpretability. This project propose a interpretable spatiotemporal deep network model, and the research objectives are three respects mainly: enhancing the ability for capturing spatiotemporal context of soil temperature, improving the interpretability of soil temperature estimation by adding the temporal interpreting and spatial interpreting components, and providing the multifarious decision-making basis for automatic meteorological observing System. By the execution of the project, the machine learning predictive model and land surface hydrological model are expected to be achieved pioneering results, and provide an important scientific value for the improvement of machine learning predictive model and the interpretability of the physical mechanism of spatial and temporal changes in soil temperature.