| Abstract Submission No. | ABS-4.1-0079 |
| Title of Abstract | High-Resolution Downscaling of Daily Minimum Temperature in India Using Advanced Deep Learning Algorithms |
| Authors | Manisha Chaturvedi and R.K. Mall* |
| Organisation | DST-MCECCR, IESD-BHU |
| Address | Lab No-123C, Second floor, IESD BHU, Varanasi, 221005 Varanasi, Uttar Pradesh, India Pincode: 221005 E-mail: manishacs13@gmail.com |
| Country | India |
| Presentation | Oral |
| Abstract | The Statistical Downscaling method serves as a robust technique for the downscaling of GCMs data to evaluate climate change at a regional level. An unprecedented increase in minimum temperature (Tmin) has been observed globally, leading to significant impacts that are particularly evident at the local level. High-resolution climate data at the local level is essential for various applications and impact assessments in agriculture, public health, environmental studies, and livelihoods. This study introduces an advanced deep learning approach: The Bidirectional Long Short-Term Memory network (BiLSTM) and the Gated Recurrent Unit (GRU) to downscale the CMIP6 GCMs models (with a coarse resolution of 1.25°) to daily minimum temperature (Tmin) data at a regional scale of 0.5° spatial high resolution for the period from 1951 to 2010 across India. The findings indicate that both methods perform exceptionally well in downscaling across all GCM model datasets; however, the BiLSTM method demonstrates slightly superior performance compared to the GRU method when evaluated against observed minimum temperatures, as well as in terms of RMSE, MAE, correlation, and spatiotemporal variability. Furthermore, the study also analyzes the future changes in Tmin over India under various future scenarios. In conclusion, the study asserts that the BiLSTM approach is a highly effective tool for downscaling (outperforming GRU) daily Tmin over India with optimal parameter tuning. Therefore, we recommend employing the BiLSTM method for downscaling GCMs datasets at a high resolution across India. |
| Keywords | CMIP6, Diurnal Temperature Range, Deep learning, Downscaling |