Deep Learning for Enhanced Multisensor Quantitative Precipitation Estimation

准确和可靠的定量降水估计(QPE)对于各种水文应用、可持续水资源规划和管理、森林火灾风险评级和农业干旱风险管理具有特别重要的意义。具体而言，高分辨率定量预测对于水文建模和预报、优化水库运行和调度、改进对山体滑坡、极端事件(洪水、干旱)的预报，包括对家庭和工业供水的季节性展望至关重要。然而，现有的QPE在精度、可靠性以及空间和时间分辨率方面都是有限的。这项拟议的研究试图通过利用人工智能、陆地表面监测技术(卫星、雷达、气象站)和气象学(改进的数值天气预报)的最新进展来解决这些限制。*这项研究计划的主要目标是开发一种创新的定量降水估计系统，以提高估计降雨量的准确性和可靠性。具体的研究目标包括：(1)研究开发QPE模型的新兴机器学习技术(即深度学习算法)；(2)开发从多个模型中合并QPE的新方法；(3)开发和评估加拿大新的多模型QPE系统。对于第一个研究目标，将确定足够的深度学习方法并用于开发有效的QPE模型。为了实现第二个目标，将基于熵理论的Copula-Bayes平均方法进行扩展，以开发更稳健的模型融合技术。最后，后者将与基于深度学习的模型一起使用，以开发多模型QPE系统。*鉴于QPE在水文学和整个水部门的重要性，以及QPE的多样化社会需求，拟议的QPE系统将极大地惠及加拿大及其他地区的广泛用户。此外，该研究计划将培训新一代高技能人员，他们将充分利用机器学习技术的最新进展，以充分利用现代观测技术提供的大数据集。

Accurate and reliable quantitative precipitation estimations (QPE) are of particular importance for diverse hydrologic applications, sustainable water resources planning and management, forest fire risk rating, and drought risk management in agriculture. Specifically, high-resolution QPE are essential for hydrologic modelling and forecasting, optimal reservoir operation and scheduling, improved forecasts of landslides, extreme events (floods, droughts), including seasonal outlooks for domestic and industrial water supply. However, existing QPE are limited in accuracy, reliability, and in their spatial and temporal resolution. The proposed research seeks to address these limitations by resorting to recent advances in artificial intelligence, land surface monitoring technology (satellites, radars, weather stations), and in meteorology (improved numerical weather predictions).******The main objective of this research program is to develop an innovative quantitative precipitation estimation system to enhance the accuracy and reliability of estimated precipitation. The specific research objectives include : (1) investigating emerging machine learning techniques (namely deep learning algorithms) for developing QPE models; (2) developing new methods for merging QPE from multiple models, (3) developing and assessing the novel multi-model QPE system across Canada. For the first research objective, adequate deep learning methods will be identified and used to develop effective QPE models. To achieve the second objective, a copula-Bayesian averaging approach will be extended based on entropy theory, to develop a more robust model fusion technique. Finally, the latter will be used with the deep learning based models to develop the multi-model QPE system.***Given the importance of QPE in hydrology and the water sector at large, and the diverse societal needs of QPE, the proposed QPE system will be highly beneficial to a wide range of users across Canada and beyond. Furthermore, the research program will train a new generation of highly skilled personnel who will be well equipped to exploit recent advances in machine learning technology to take full advantage of the large datasets been made available by modern observing technology.