Specialised Environments

Page tree

Background

Inferring local and regional climate information from global climate models is a regular task in Earth and Climate sciences. The process of obtaining high-resolution local climate information from coarse-resolution global models is known as downscaling. All the various techniques for downscaling can be put into two broad categories, namely dynamical and statistical. Dynamical downscaling uses numerical models to produce high-resolution local information and is computationally expensive. On the other hand, statistical downscaling uses a variety of statistical patterns to establish a relationship between a coarse global climate model and observed local climate data.


One of the benefits of statistical downscaling is that Deep Learning models can be used to learn statistical patterns. In recent years several Deep Learning models have been put forward for statistical downscaling based on various convolution and super-resolution architectures. Not every model produces the best result in every situation; therefore, a framework is required to compare the performance of various architectures.

DL4DS

DL4DS (Deep Learning for Empirical Downscaling) is based on the following paper: https://arxiv.org/pdf/2205.08967.pdf. It is a Python library that provides a common interface to work with several different deep learning architectures. It is highly configurable and allows users to choose between several learning strategies. It has multi-node, multi-GPU support, enabling users to perform comparative studies on large-scale data. In this page, first, we demonstrate how the DL4DS models can be configured to train with the CORDEX data, and then some of the results are briefly discussed.

Architecture

Figure 1 shows the architecture of the DL4DS framework. On the top right, components of the input training data are shown, it consists of high and low-resolution samples and optional static and predictor variables. The high and low-resolution data are fed in pairs and contain the patterns that the deep learning models suppose to learn. The static and predictor variables are optional and can provide spatiotemporal information about the region like the orographic or land-sea mask data that can help the model. 

On the right-hand side, one can see that the input batches are prepared from high-resolution, low-resolution, and options data, and then the data is fed to the model. As mentioned earlier the DL4DS has several options for models. At present, DL4DS offers four different backbones training and for each different modules can be selected, they are discussed next.

Figure 1. DL4DS Architecture.


Modules and Backbones

DL4DS provides several pre-built modules and backbones, and they can be used to build new models. Figure 2 shows the available model building blocks, on the left-hand side each block is named and presented in a separate colour. They represent the fundamental layers of the deep learning network. On the right-hand side, the conceptual view is presented how the modules can be arranged to build various models. Those modules are available as APIs from the DL4DS library.

Figure 2. Building blocks of DL4DS.


In addition, DL4DS also provides different backbones that can be used for downscaling of parameters. The deep learning backbones provided by DL4DS include Convnet, Resnet, Densenet, Unet, and ConvNext. Figure 3 provides some conceptual views of how to organize the above-mentioned building blocks. Figure 3 (A) shows the layers are arranged with a skip connection, this architecture is used for Resnet and Densenet. Figure 3 (B) shows the blocks arranged for the ConvNext model. Figure 3 (C) shows the encoder-decoder architecture of the Unet. Figure 3 (D) shows the architecture of the recurrent-convolution network. 

Lastly, Figure 3 (E) shows how the final output is constructed from the backbone output. Training data is passed through any of the four above-mentioned backbones (A-D), afterwards, the backbone output is fed through localized convolution layers and concatenated with optional static variables. Finally, a series of transition and spatial convolution layers ensembles data from all the components to produce the high-resolution output.  

Figure 3. DL4DS deep learning models.

NCI DL4DS examples

You may find more tutorial and projects about running DL4DS at NCI.




































































































































  • No labels