基于可解释机器学习模型的全球月天然径流估算方法研究

Natural flow estimation is the basis for the study of natural flow change and its driving force, which is of great significance for water resources planning, river protection and ecological environment restoration. This study introduces the spatial and temporal heterogeneity factors and the anthropogenic disturbance factors. A large-scale monthly natural flow estimation model based on a intelligible machine learning is proposed by using global remote sensing and streamflow data. We evaluate the regional sensitivity and uncertainty of the model. After that, we analyze the importance of runoff impact factors which was evaluated in the model. Finally we reversely explore the mechanisms of how the impact factors impact natural flow. This study provides a new idea for monthly natural flow estimation in large-scale ungauged areas, and provides a new understanding of the mechanism of runoff generation.

天然径流估算是开展径流情势变化及其驱动力研究的基础，对水资源规划、河流保护与生态环境修复都具有重要意义。本项目利用全球遥感和径流资料，引入时空异质因子、人类活动扰动因子，评估“参考站”受人类活动的影响，构建基于可解释机器学习的大尺度月天然径流估算模型；评估机器学习模型的区域敏感性和不确定性；根据模型结果，分析径流影响因子的重要性，逆向探索影响因子作用机理。本研究为大尺度无资料地区月天然径流估算提供新思路，为径流形成机理提供新认识。

天然径流估算是开展径流情势变化及其驱动力研究的基础，对水资源规划、河流保护与生态环境修复都具有重要意义。因此，本项目使用全球径流观测资料和多源遥感等产品，构建了基于机器学习方法的全球天然径流估算大数据框架，取得了比传统水文模型更好的结果，发展了天然径流估算方法；实现了对全球和地区天然径流的高精度估算，生成了一套具有较高精度的高分辨率天然径流空间分布数据集。研究成果为全球和地区的水资源评价、水资源开发利用评价、水资源安全评价等提供数据支撑和方法基础。