RainNet

人工智能科学方法已被应用于解决科学问题 (例如核聚变，生态学，基因组学，气象学)，并取得了非常有希望的结果。空间降水降尺度是最重要的气象问题之一，迫切需要AI的参与。但是，缺乏组织良好且带有注释的大规模数据集，阻碍了对更有效且先进的深度学习模型进行降水降尺度的培训和验证。为了缓解这些障碍，我们提出了第一个名为RainNet的大规模空间降水缩小数据集，其中包含超过17年的62,400对高质量低/高分辨率降水图，准备帮助深度学习模型在降水缩小中的演变。具体而言，在雨网中仔细收集的降水图涵盖了各种气象现象 (例如飓风，狂风)，这对提高模型泛化能力有很大帮助。此外，雨网中的地图对以图像序列的形式组织 (每月720地图或1地图/小时)，显示复杂的物理特性，例如时间未对准、时间稀疏和流体特性。专门引入了两个面向深度学习的度量标准，以评估或验证训练模型的综合性能 (例如，预测图重建精度)。为了说明RainNet的应用，评估了14种最先进的模型，包括深层模型和传统方法。为了充分探索潜在的降尺度解决方案，我们提出了一个隐式物理估计基准框架来学习上述特征。广泛的实验证明了雨网在训练和评估降尺度模型中的价值。

Artificial intelligence (AI)-enabled scientific methods have been applied to address scientific problems (e.g., nuclear fusion, ecology, genomics, meteorology), yielding highly promising results. Spatial precipitation downscaling is one of the most critical meteorological issues, and the involvement of AI is urgently needed. However, the lack of well-organized, annotated large-scale datasets has hindered the training and validation of more efficient and advanced deep learning models for precipitation downscaling. To alleviate these barriers, we present RainNet, the first large-scale spatial precipitation downscaling dataset, which contains 62,400 pairs of high-quality low- and high-resolution precipitation maps spanning over 17 years, intended to support the development of deep learning models for precipitation downscaling. Specifically, the precipitation maps carefully collected in RainNet cover a wide range of meteorological phenomena (e.g., hurricanes, gales), which greatly contributes to improving the generalization ability of models. In addition, the maps in RainNet are organized as image sequences (720 maps per month or 1 map per hour), exhibiting complex physical characteristics such as temporal misalignment, temporal sparsity, and fluid dynamical properties. Two deep learning-oriented metrics are specifically introduced to evaluate or validate the comprehensive performance of trained models, such as the reconstruction accuracy of predicted precipitation maps. To illustrate the applicability of RainNet, we evaluate 14 state-of-the-art models spanning deep learning-based approaches and traditional methods. To fully explore potential downscaling solutions, we propose an implicit physical estimation benchmark framework for learning the aforementioned characteristics. Extensive experimental results validate the value of RainNet in training and evaluating precipitation downscaling models.

搜集汇总

背景与挑战

背景概述

RainNet是一个大规模空间降水降尺度数据集，包含62,400对高质量降水图，覆盖17年数据，支持深度学习模型训练和验证。数据集特点包括多样的气象现象、复杂的物理特性和专门的评估指标，旨在推动降水降尺度领域的研究。