Deep Learning-based Non-Stationary Bias Correction (NSBC)

Abstract

This work develops the NSBC (non-stationary bias correction) methodology to correct temperature projection bias from E3SM. The NSBC deep learning framework consists of a three-part architecture: an auto-encoder for compressing the spatial information, an LSTM for predicting annual temperature mean, and a U-Net for capturing the residual bias in temperature. The non-stationary bias correction (NSBC) framework can correct the non-stationarity of the biases of the climate models, which significantly improves the accuracy of future temperature prediction and improves the overestimation of extreme high temperatures that many existing bias correction methods suffer from. Getting started 1. Obtain the historical climate simulation and observation data. The E3SM simulation data are available through https://aims2.llnl.gov/search/cmip6/. The pseudo observations, the Geophysical Fluid Dynamics Laboratory (GFDL)-ESM4 model (Krasting et al., 2018) are available through https://aims2.llnl.gov/search/cmip6/. The spatial resolution of E3SM and pseudo observation datasets are both regridded to a common 1° resolution grid using conservative interpolation. The regridded E3SM and pseudo observation with 1° resolution can be found throught ./data/. 2. Train the Auto-encoder model. Python 0-autoencoder.py 3. Train the LSTM Python 1-LSTM.py 4. Generate the annual mean temperature based on trained LSTM Python 2-generate_annual_mean_LSTM.py 5. Train the U-Net. Python 3-unet.py 6. Evaluation and compared with the baseline Python 4_evaluation.py Is there a deadline approaching that requires the release of yo