Augmented Convolutional LSTMs for Generation of High-Resolution Climate Change Projections

Abstract

Projection of changes in extreme indices of climate variables such as temperature and precipitation are critical to assess the potential impacts of climate change on human-made and natural systems, including critical infrastructures and ecosystems. While impact assessment and adaptation planning rely on high-resolution projections (typically in the order of a few kilometers), state-of-the-art Earth System Models (ESMs) are available at spatial resolutions of few hundreds of kilometers. Current solutions to obtain highresolution projections of ESMs include downscaling approaches that consider the information at a coarsescale to make predictions at local scales. Complex and non-linear interdependence among local climate variables (e.g., temperature and precipitation) and large-scale predictors (e.g., pressure fields) motivate the use of neural network-based super-resolution architectures. In this work, we present auxiliary variables informed spatio-temporal neural architecture for statistical downscaling. The current study performs daily downscaling of precipitation variable from an ESM output at 1.15 degrees (115 km) to ¼ degrees (25 km) over the one of the most climatically diversified countries, India. We showcase significant improvement gain against two popular state-of-the-art baselines with a better ability to predict statistics of extreme events. To facilitate reproducible research, we make available all the codes, processed datasets, and trained models in the public domain.

Depth‐Duration‐Frequency of Extreme Precipitation Events Under Internal Climate Variability: Indian Summer Monsoon

“…The main objective is to adjust the statistical properties of climate simulations with observations. Several studies have used QM bias correction for annual precipitation maxima (Cannon et al., 2015; Harilal et al., 2021; Maraun, 2013; Ngai et al., 2017). The QM has been shown to be an effective method for removing biases at relatively little computational expense (Li et al., 2010; Maraun, 2013; Maraun et al., 2010; Maurer & Pierce, 2014).…”

DL4DS -- Deep Learning for empirical DownScaling

“…Statistical downscaling of gridded climate variables is a task closely related to that of superresolution in computer vision, considering that both aim to learn a mapping between lowresolution and high-resolution grids (Wang et al, 2021). Unsurprisingly, several DL-based approaches have been proposed for statistical or empirical downscaling of climate data in recent years (Vandal et al, 2017;Leinonen et al, 2020;Stengel et al, 2020;Höhlein et al, 2020;Liu et al, 2020;Harilal et al, 2021). Most of these methods have in common the use of convolutions for the exploitation of multivariate spatial or spatio-temporal gridded data, that is 3D (height/latitude, width/longitude, channel/variable) or 4D (time, height/latitude, width/longitude, channel/variable) tensors.…”

“…The trained model is then applied to the desired, unseen during training, lowresolution data in a domain-transfer fashion. Other approaches such as those proposed by Höhlein et al (2020) or Harilal et al (2021), model a cross-scale transfer function between explicit low-resolution and high-resolution datasets. DL4DS supports both training frameworks, with explicit paired samples (in MOS fashion) or with paired samples simulated from high-resolution fields (in PerfectProg fashion).…”

“…Convolutional Long Short-Term Memory (ConvLSTM, Shi et al, 2015) and convolutional Gated-Recurrent-Units (Ballas et al, 2015) are examples of models that have been used for downscaling time-evolving gridded data (Leinonen et al, 2020;Harilal et al, 2021). DL4DS can model either spatial or spatio-temporal data, by either using standard convolutional blocks or recurrent convolutional blocks.…”

“…In the context of empirical downscaling, pre-upsampling has been used by Vandal et al (2017); Harilal et al (2021), resize convolution by Leinonen et al (2020), transposed convolution by Höhlein et al (2020) and sub-pixel convolution by Liu et al (2020). Choosing an upsampling method depends on the problem at hand (MOS or PerfectProg training) and the computational resources available.…”