2022-06-20, 11:15–11:45, Room 2
In this talk, we present DL4DS, a python package that implements a wide variety of state-of-the-art and novel algorithms for downscaling gridded Earth Science data with deep neural networks. DL4DS has been designed with the goal of providing a general framework for convolutional neural networks with configurable architectures and training procedures to enable benchmark, comparative and ablation studies.
A common task in Earth Sciences (ES) is to infer weather and climate information at local and regional scales from global climate models. Dynamical downscaling requieres running expensive numerical models at high resolution and comes with high computational costs. Empirical or statistical downscaling techniques present an alternative approach for learning links between the large- and local-scale climate. A large number of deep neural network-based approaches for statistical downscaling have been proposed in recent years (Vandal et al., 2017; Leinonen et al., 2020; Höhlein et al., 2020; Liu et al., 2020; Harilal et al., 2021), mainly in the form of convolutional neural networks (CNNs). CNNs have been used almost universally for the past few years in computer vision applications, such as object detection, semantic segmentation and super-resolution (SR), and have shown outstanding capabilities in the task of SR and downscaling Earth observation and ES data.
While scientific software tools such as Numpy, Xarray or Jupyter have an essential role in modern ES research workflows, state-of-the-art domain-specific Deep Learning-based algorithms are usually developed as proof-of-concept scripts. For Deep Learning to fulfil its potential to transform ES, the development of Artificial Intelligence- and Deep Learning-powered scientific software must be carried out in a collaborative and robust way following open-source and modern software development principles. In our search for efficient architectures for downscaling ES data, we have developed DL4DS, a library that draws from recent computer vision developments for tasks such as image-to-image translation and SR. DL4DS is implemented in Tensorflow/Keras and contains a large collection of building blocks that abstract and modularize a few core design principles for composing and training CNNs-based empirical downscaling models. DL4DS can be found in this repository: https://github.com/carlgogo/dl4ds