Spatio-Temporal Deep Learning for Rapid Time-Series Forecasting and Data Synthesis

许多科学和工程领域都需要进行时空分析，如用于预警的洪水风险预测、用于空气质量区管理的空气污染预测、用于精确草坪管理等。目前的模型采用复杂的模型来估计必要的时空输入。尽管这些方法是成功的，但它们的计算量很大，由于运行时间长、数据集不可用或过时，无法为本地化需求提供结果。缺乏准确和最新的模型数据集会极大地影响结果和基于这些结果的决策，特别是在应急管理情况下。为了解决现有时空方法的局限性，该项目将开发新的基于深度学习的架构，以便在空间和时间上快速评估复杂和时间敏感的管理问题。本研究的结果将解决时间序列估计（TSE）模型需要的局部时空需求，通过更新/填补时间和空间输入的数据空白，以快速构建更准确的预测估计。由此产生的方法将为实时预报提供近乎瞬时的模型输入更新，采用遥感，整合来自分布式传感器网络、监测站和基于网络的资源等的测量。我们的目标是能够获取当前的图像和详细的地理空间数据集，并将它们转换成可用于精确建模时空事件的格式，例如洪水和空气浓度随时间的分散。这些事件直接受到地形、土地利用和当地自然特征的影响。通过提供局部区域的最新表示，模型输出将越准确。这些新方法将提供接近实时的信息，以支持应急响应人员、环境规划者和其他人员

There are many scientific and engineering fields that require spatio-temporal analysis such as flood risk forecasting for early warning, air pollution forecasting for air quality zone management, and precision sod management. Current models employ complex models to estimate necessary spatio-temporal inputs. Albeit successful, these methods are computationally intensive and cannot provide results for localized requirements due to long run times and unavailable or out-of-date datasets. The lack of accurate and up to date datasets for models can greatly impact results and decisions based on these results, especially in emergency management situations. To address the limitations of existing spatio-temporal methods, this project will develop new deep learning-based architectures to allow for the rapid evaluation of complex and time-sensitive management problems in space and time. The results of this research will address the local spatio-temporal requirements that time-series estimation (TSE) models need by updating/filling data gaps in temporal and spatial inputs to rapidly construct more accurate forecast estimates. The resulting methodology will provide near-instantaneous model input updates for real-time forecasts, employing remote sensing, integrating measurements from networks of distributed sensors, monitoring stations, and web-based resources, among others. Our goal is to be able to take current imagery and detailed geospatial datasets and convert them into formats that can be used to accurately model spatio-temporal events such as flooding and air concentration dispersion over time. These events are impacted directly by topography, land use, and local physical features. By providing the most up to date representation of the local area, the more accurate the model outputs will be. These new methodologies will provide near-real-time information to support emergency responders, environmental planners, and other decision-makers.