PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepSD: Generating High Resolution Climate Change Projections through Single Image Super-Resolution

Thomas Vandal, Evan Kodra, Sangram Ganguly, Andrew Michaelis, Ramakrishna Nemani, Auroop R Ganguly

arXiv:1703.03126v1cs.CV

TL;DR

Coarse ESMs and uneven observations complicate localized climate projections, while existing downscaling methods can struggle with extremes. DeepSD adapts stacked, multi-scale super-resolution convolutional networks to statistical downscaling and outperforms compared approaches on reported daily precipitation metrics while enabling scalable ensemble processing. The paper also identifies untested spatial and temporal generalization as limitations.

  • Problem

    Coarse ESM resolution, sparse observations, and existing methods’ difficulty with extreme events limit reliable localized climate projections.

  • Method

    DeepSD uses an augmented stacked super-resolution convolutional neural network with multi-scale information, observations, and topography for statistical downscaling.

  • Results

    DeepSD outperforms the other tested approaches in bias, correlation, and RMSE, has lower RMSE than BCSD for 79% of CONUS, and supports ensemble downscaling.

  • Takeaways & Limitations

    Spatial information from super-resolution networks may support downscaling beyond grid-by-grid predictions and scalable generation of products from diverse ESM ensembles.

  • Takeaways & Limitations

    The experiments do not test predictability in untrained regions or temporal non-stationarity.

Abstract

from arXiv · show

The impacts of climate change are felt by most critical systems, such as infrastructure, ecological systems, and power-plants. However, contemporary Earth System Models (ESM) are run at spatial resolutions too coarse for assessing effects this localized. Local scale projections can be obtained using statistical downscaling, a technique which uses historical climate observations to learn a low-resolution to high-resolution mapping. Depending on statistical modeling choices, downscaled projections have been shown to vary significantly terms of accuracy and reliability. The spatio-temporal nature of the climate system motivates the adaptation of super-resolution image processing techniques to statistical downscaling. In our work, we present DeepSD, a generalized stacked super resolution convolutional neural network (SRCNN) framework for statistical downscaling of climate variables. DeepSD augments SRCNN with multi-scale input channels to maximize predictability in statistical downscaling. We provide a comparison with Bias Correction Spatial Disaggregation as well as three Automated-Statistical Downscaling approaches in downscaling daily precipitation from 1 degree (~100km) to 1/8 degrees (~12.5km) over the Continental United States. Furthermore, a framework using the NASA Earth Exchange (NEX) platform is discussed for downscaling more than 20 ESM models with multiple emission scenarios.

1 INTRODUCTION

Climate change threatens infrastructure, ecosystems, resources, and vulnerable populations, while coarse ESM outputs limit localized impact assessment. DeepSD applies spatial super-resolution to statistical downscaling and supports daily, multi-scenario ensemble projections through NASA’s Earth Exchange.

  • Climate change intensifies extreme events and threatens natural resources, transportation, energy systems, infrastructure, health, and security.
  • ESMs provide physically based climate projections but their coarse spatial resolution limits localized impact assessment.
  • Statistical downscaling learns mappings from low-resolution ESM outputs to high-resolution projections using observations, but sparse observations leave some regions without needed data.
  • DeepSD adapts super-resolution convolutional neural networks to capture spatial information and improve beyond existing downscaling methods.
  • DeepSD supports scalable daily downscaling of ensemble ESMs over CONUS for four emission scenarios using NASA’s Earth Exchange platform.
  • The paper compares DeepSD with BCSD and automated-statistical downscaling methods, while positioning improved daily products as useful for studying local climate impacts.

2 EARTH SCIENCE DATA

Earth science datasets combine rich spatial and temporal information with substantial computational demands, while ESM outputs remain limited by resolution and precipitation disagreement. Observations and topography provide complementary information for learning spatially structured downscaling models.

  • Daily 4km data for one climate variable over the United States requires 1.2GB of storage, illustrating earth science data’s computational and storage burden.
  • ESMs couple atmosphere, land, and ocean processes but exhibit low resolution and limited model agreement, particularly for precipitation.
  • Statistical downscaling learns mappings from ESM outputs to higher-resolution observations sourced from satellites, stations, or reanalysis datasets.
  • Topography affects weather and climate patterns, with different spatial scales capturing large-scale weather patterns and fine-resolution precipitation biases.
  • Earth science products contain spatial dependencies analogous to image channels, but traditional regression methods often remove this structure by vectorizing spatial data.

3 STATISTICAL DOWNSCALING

Statistical downscaling maps low-resolution climate variables to high-resolution projections using observations, but the mapping is ill-posed and uncertainty is amplified by climate non-stationarity and precipitation sparsity.

  • Statistical downscaling maps a low-resolution climate variable to a high-resolution projection, transforming one grid point into multiple points.
  • The mapping is ill-posed because many high-resolution solutions can correspond to one low-resolution input.
  • High-resolution topography and other low-resolution climate variables can mitigate the ill-posed problem by adding spatial information.
  • Observed climate datasets train the statistical model, which then infers downscaled ESM projections.
  • Spatial and temporal non-stationarity, sparse precipitation occurrences, and skewed rainfall distributions propagate uncertainty in downscaled projections.

4 RELATED WORK

Related downscaling methods include regression, weather classification, and BCSD, but many perform poorly for extremes or discard spatial structure. DeepSD draws on image super-resolution and convolutional neural networks to address this gap.

  • Statistical downscaling includes regression, weather classification, and weather generators, with regression and classification reviewed for spatial resolution increases.
  • Automated Statistical Downscaling classifies precipitation days before regressing rainfall amounts, while BCSD bias-corrects, interpolates, and spatially scales ESM outputs.
  • Weather classification methods select analogous high-resolution observations using distances between ESM and observed low-resolution conditions.
  • Existing approaches often estimate mean precipitation adequately but tend to fail when downscaling extreme events.
  • Few prior studies explicitly model spatial properties in downscaled projections, motivating an image analogy while recognizing that climate data and images are not directly equivalent.
  • Single-image super-resolution learns mappings between low- and high-resolution image pairs, paralleling the statistical downscaling task.
  • Patch-based super-resolution, dictionary learning, kernel regression, random forests, and neighborhood regression parallel methods used in statistical downscaling.
  • Convolutional neural networks generalize sparse coding through nonlinearity, flexible architectures, and scalability to large datasets.

5 METHODOLOGY

DeepSD adapts SRCNNs into a stacked, multi-scale architecture for statistical downscaling, using successive resolution enhancements and high-resolution topographical inputs. Each SRCNN learns an independent low-resolution-to-high-resolution mapping before the networks are stacked at test time.

  • Super-resolution CNN: SRCNN learns a low-resolution-to-high-resolution mapping through patch extraction, nonlinear mappings, and reconstruction.The network uses a three-layer CNN and Euclidean loss over training inputs and high-resolution labels.
  • Super-resolution CNN: Layer 1 extracts high-dimensional patch vectors, layer 2 applies nonlinear mappings, and layer 3 reconstructs the high-resolution estimate.The first layer uses overlapping filters, while the final convolution produces the reconstructed image.
  • Stacked SRCNN: Stacked SRCNNs increase resolution successively by a factor of s, enabling large 8–12x enhancements through multi-scale spatial representations.Each SRCNN is trained independently for its input and output resolutions, then stacked at test time.
  • Stacked SRCNN: Cascading SRCNNs performed worse than stacked SRCNNs in the authors’ experiments, potentially because cascading propagates errors through successive networks.Stacked training uses independently generated input/output pairs at multiple scales.
  • DeepSD: DeepSD augments stacked SRCNNs with precipitation and high-resolution topographical channels, including elevation and land/water information.Static topographical features can be used at each scaling factor because they are known at high resolution and generally remain unchanged.
  • DeepSD: The stacked process allows DeepSD to capture both regional and local patterns.The architecture is presented as an augmented stacked super-resolution convolutional neural network for statistical downscaling.

6 APPLICATION OF DEEPSD

The application downscales daily CONUS precipitation using PRISM precipitation and GTOPO30 elevation, targeting an eightfold increase from 1.0° to 1/8° resolution. DeepSD uses three successive twofold SRCNN enhancements with precipitation and elevation as inputs.

  • Motivation and data: High-resolution precipitation is important for climate adaptation, while existing downscaling methods may add little information beyond coarse ESM precipitation.The application therefore focuses on daily precipitation over the data-rich Continental United States.
  • Resolution enhancement: DeepSD increases resolution eightfold from 1.0° to 1/8° through three successive 2x SRCNN stages.The sequence is 1.0°→1/2°→1/4°→1/8°.
  • Data representation: Each daily 1/8° CONUS image has size 208x464, with precipitation and elevation as inputs and precipitation as the sole output.Lower-resolution inputs are produced by bicubic up-sampling and interpolation for the corresponding SRCNN stage.

Data preprocessing.

Training data are prepared as multiscale sub-images, with normalization and progressively more samples at higher resolutions. The SRCNN stages share an architecture and are optimized with Adam using layer-specific learning rates.

  • Sub-image preparation: Training extracts 51x51 sub-images at stride 20 to increase heterogeneity, with sample counts rising from 1,095 to 45,625 per year across resolutions.Features and labels are normalized to zero mean and unit variance.
  • Network parameters: All SRCNNs use 64 9x9 filters in layer 1, 32 1x1 filters in layer 2, and a 5x5 output kernel.The paper notes that higher-resolution models may benefit from larger kernels and more filters.
  • Optimization: Adam optimization uses learning rates of 10^-4 for the first two layers and 10^-5 for the last layers.The same parameter set is used across the SRCNNs.

6.2 Comparison

The comparison evaluates DeepSD against BCSD and three Automated-Statistical Downscaling methods using predictive metrics across locations, seasons, and precipitation extremes. DeepSD generally provides stronger predictive performance than the baselines, while BCSD remains strong for distributional agreement and runtime.

  • Baseline Method: BCSD uses interpolation and training-set scaling factors to reconstruct high-resolution precipitation from low-resolution inputs.The implementation uses 1980–2005 for training and 2006–2014 for testing.
  • Baseline Methods: Automated-Statistical Downscaling separates precipitation downscaling into rainy-day classification and precipitation estimation, using logistic/lasso regression, SVM, or ANN models.The compared ASD approaches combine classification and regression components.
  • Daily Predictability: DeepSD outperforms BCSD, Lasso, SVM, and ANN in bias, correlation, and RMSE, while closely trailing BCSD in skill.The comparison averages four metrics over 1000 randomly selected CONUS locations.
  • Method Selection: DeepSD and BCSD are retained for subsequent experiments because they are the two scalable, top-performing methods from the initial comparison.BCSD performs well for minimizing bias and estimating the underlying precipitation distribution.
  • Seasonal Comparison: DeepSD has higher correlation and lower RMSE than BCSD across seasons, with lower RMSE at 79% of CONUS locations.Both methods achieve low bias below 0.5 mm/day and skill of approximately 0.98.
  • Extreme Precipitation: The extreme-precipitation experiment compares DeepSD and BCSD over percentile thresholds from 90 to 99.9 using RMSE, correlation, bias, and skill.Metrics are computed for events exceeding each threshold and averaged over CONUS.

6.3 Scalability on NASA’s NEX

The paper assesses computational scalability for downscaling the multi-model, multi-scenario CMIP5 ensemble. DeepSD can generate large CONUS datasets with GPU, MPI, and NASA infrastructure, although the workload still requires substantial computational resources.

  • Ensemble Scope: CMIP5 includes more than 20 models, four emission scenarios, multiple initial conditions, and multiple daily climate variables.The ensemble is used to represent uncertainty across projections.
  • Runtime Evaluation: DeepSD runtime is evaluated for downscaling one year of CONUS data using three feed-forward processes and interpolations on a single GPU.ASD runtimes are scaled from 1000 locations processed with 40 CPUs in parallel, while BCSD runtime is measured separately.
  • Scalability: BCSD is the quickest method, but DeepSD remains scalable through a single feed-forward neural-network architecture.The authors report that generating very large datasets still requires heavy computational power.
  • NASA NEX Infrastructure: NASA’s Pleiades supercomputer, GPU-enabled nodes, MPI, and NEX storage support rapid computation and storage of high-resolution projections.NEX provides 2.3PB of rapid-access storage plus large-scale tape storage.

7 CONCLUSION

DeepSD shows promise as a scalable statistical-downscaling architecture, but its experimentation leaves important generalization and uncertainty questions unresolved.

  • Spatial and temporal generalization: DeepSD’s experimentation does not test predictability in regions where the model was not trained.The authors identify this as an unresolved question for assessing credibility in regions with few observations.
  • Spatial and temporal generalization: The study does not test temporal non-stationarity, a longstanding problem in statistical downscaling.The authors propose splitting training and testing data between cold and warm years for evaluation.
  • Future improvements: DeepSD currently fails to capture uncertainty around its downscaled projections.The paper identifies Bayesian deep learning as a possible direction for quantifying uncertainty.
  • Future improvements: Future work could add variables such as temperature, wind, and humidity and jointly downscale multiple climate variables.These extensions may help capture additional climate patterns and relationships in high-resolution datasets.
  • Conclusion: Despite these limitations, DeepSD is presented as a scalable architecture with high predictive capabilities and a novel framework for statistical downscaling.
Loading 1703.03126v1…