Reconstructing High Resolution ESM Data Through a Novel Fast Super Resolution Convolutional Neural Network (FSRCNN)

Abstract

Accurate and reliable climate data is critical for assessing the risk of climate change to our society's well-being. Increases in temperature, sea-level, and extreme weather events can render many aspects of our society vulnerable including our health, natural resources, and energy-systems (Nicholls & Cazenave, 2010;Trenberth, 2012). Local and regional climate future projection data is the most crucial for planning and mitigating these risks, but is often the least reliable (Schmidt, 2010). Currently, Earth System Models (ESMs) used for simulating Earth's past climate and future projections are often computationally limited to coarse horizontal resolutions, generally between 1° and 3° (Vandal et al., 2017). These low resolution models fail to accurately simulate important physical processes such as precipitation extremes (Kharin et al., 2007). Recent advances in computing resources have allowed for global ESMs to be run at higher resolutions (∼0.25°) for longer time periods and have been shown to improve the simulations of regional mean climate as well as extremes (Mahajan et al., 2015;Wehner et al., 2010). However, these high resolution models remain prohibitively expensive.A computationally less expensive approach to derive high resolution climate data over a region of interest is to map data from low resolution global model simulations to high resolution grids using dynamic or statistical downscaling techniques. Dynamical downscaling involves running high resolution regional dynamical models to extrapolate large scale boundary conditions obtained from a coarser global ESM to finer resolutions on regional scales. Statistical downscaling (SD) aims to map coarse resolution data to high resolution projections using statistical methods like linear regression. Recent studies have shown that machine learning techniques, like neural networks (Fistikoglu & Okkan, 2011;Vu et al., 2016) and support vector machines (Ghosh, 2010), for SD significantly outperform other traditional SD methods. In this study, we use one such computer vision approach called super-resolution (SR), which generates a high resolution image from its low resolution equivalent. SR techniques attempt to generalize across images and have been shown to learn local scale patterns more efficiently than other downscaling methods (Vandal et al., 2017).

On the limitations of deep learning for statistical downscaling of climate change projections: The transferability and the extrapolation issues

“…Convolutional Neural Networks (CNNs) can deal with large amounts of data and they present an important advantage over other statistical methods; their ability to extract high‐level spatial features automatically (LeCun et al, 1998; LeCun & Bengio, 1995). CNNs have become one of the main state‐of‐the‐art downscaling techniques, both as Perfect Prognosis SD methods and as hybrid approaches (see Baño‐Medina et al, 2020, 2021; Höhlein et al, 2020; Liu et al, 2023; Passarella et al, 2022; Serifi et al, 2021; Vandal et al, 2017, 2019). The particular implementation UNET (Ronneberger et al, 2015) has been widely used for image recognition with great performance and different variations have been also satisfactorily applied to climate downscaling (Doury et al, 2023; Sha et al, 2020a, 2020b; Sharma & Mitra, 2022).…”

Using Explainability to Inform Statistical Downscaling Based on Deep Learning Beyond Standard Validation Approaches

“…CNNs exhibit the ability to automatically infer spatial features that encode predictive information from the predictor fields. The most popular application of deep learning for downscaling follows the Super Resolution (SR) approach (Vandal et al., 2018; Passarella et al., 2022; J. Wang et al., 2021; Vandal et al., 2018), inspired by the homonym field of computer vision (Z. Wang et al., 2020). This approach learns to generate high‐resolution fields (predictand) from their low‐resolution counterparts; therefore, it is not well‐suited for GCM downscaling under the PP approach due to the aforementioned limitations of surface predictors.…”

“…CNNs exhibit the ability to automatically infer spatial features that encode predictive information from the predictor fields. The most popular application of deep learning for downscaling follows the Super Resolution (SR) approach (Vandal et al, 2018;Passarella et al, 2022;J. Wang et al, 2021;Vandal et al, 2018), inspired by the homonym field of computer vision (Z.…”

Improved Generative Adversarial Network for Super-Resolution Reconstruction of Coal Photomicrographs

“…Very Deep Super-Resolution (VDSR), with 20 residual layers, enhances the performance of super-resolution image reconstruction, whereas it consumes much more computational cost [ 11 , 12 ]. Fast Super-Resolution Convolutional Neural Network (FSRCNN) has a relatively shallow network structure consisting of four convolution layers and one deconvolution layer [ 13 ]. It was demonstrated to have faster speed and better reconstructed image quality than the SRCNN.…”