Deep Learning Atmospheric Features

在气象数据中探测有趣的特征（气旋、大气河流、锋面）有着不断提高保真度的悠久历史，因为开发了利用越来越多的计算机能力和更复杂的数学的新技术。最近，通过增加从计算机视觉借来的更新技术，工具库已经增加。早期的工作表明，它们在特征检测方面可能比现有技术具有更高的准确性，但在深度学习的情况下，这对计算提出了巨大的要求-至少在应用于原始数据时-以及对数据存储的巨大要求。（即，在一组特定场景下可能发生的许多实现）。今天，我们将写出所有数据，然后研究输出数据中有趣的特征。随着分辨率的提高，仅仅存储所有数据将是困难的，更不用说对存储的数据进行特征检测了。未来处理这个问题的一个可能的方法是在模型运行时寻找每个集合成员的特征--只写出我们已经知道具有有趣特征的模拟。这个项目是关于测试和比较一系列检测技术对原始高分辨率气候模型数据和相同数据的降低精度和降低分辨率变体的保真度。我们的目标是开发一种学习技术，可以在模型模拟过程中使用，并识别这些特征，而无需花费大量的计算成本。这很可能是通过使用数据的简化版本来完成的，而不是改变技术，但是不同的技术可能在不同级别的简化数据上工作得更好。然而，如果我们确实需要改变和/或修改技术，我们会问自己：“对于一个强大而多功能的学习来说，哪些是最重要的基石？哪些是有效训练的关键内层和神经元？"

The detection of interesting features (cyclones, atmospheric rivers, fronts) in meteorological data has a long history of ever-increasing fidelity as newer techniques have been developed exploiting increasing computer power and more mathematical sophistication. Recently, the armoury of tools has been increased by the addition of newer techniques borrowed from Computer Vision. Early work suggests they may have higher levels of accuracy in feature detection than existing techniques, but in the case of Deep Learning this comes with enormous requirements on computing - at least when applied to the raw data - and with that enormous requirements on data storage.In the future we expect to run large ensembles of climate simulations (that is, many realisations of what might occur under a specific set of scenarios). Today we would write out all the data, and then investigate the output data for interesting features. As resolution increases, just storing all that data will be difficult, let alone doing the feature detection on the stored data. A possible way to handle this in the future would be to look for the features in each one of the ensemble members as the model runs - and only write out the simulations which we already know have interesting features.This project is about testing and comparing the fidelity of a range of detection techniques to both raw high- resolution climate model data and both reduced precision and reduced resolution variants of the same data. The goal will be to develop a learning technique that can be used during the model simulation and identify such features without itself being enormously computationally expensive. This will most likely be done by using reduced versions of the data, rather than changing the techniques, but different techniques may work better at different levels of reduced data. However, if we do need to change and/or modify techniques, we will be asking ourselves "Which are the most important building blocks for a robust and versatile learning? Which are the crucial inner layers and neurons for effective training?"