Japan Geoscience Union Meeting 2023

Presentation information

Print

[E] Oral

A (Atmospheric and Hydrospheric Sciences ) » A-TT Technology &Techniques

[A-TT29] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

Mon. May 22, 2023 1:45 PM - 3:00 PM Exhibition Hall Special Setting (4) (Exhibition Hall 8, Makuhari Messe)

convener:Venkata Ratnam Jayanthi(Application Laboratory, JAMSTEC), Patrick Martineau(Japan Agency for Marine-Earth Science and Technology), Takeshi Doi(JAMSTEC), Swadhin Behera(Application Laboratory, JAMSTEC, 3173-25 Showa-machi, Yokohama 236-0001), Chairperson:Patrick Martineau(Japan Agency for Marine-Earth Science and Technology), Takeshi Doi(JAMSTEC), Venkata Ratnam Jayanthi(Application Laboratory, JAMSTEC)

2:15 PM - 2:30 PM

[ATT29-03] India summer monsoon rainfall prediction over core monsoon region at short to long leads using deep learning

*Kalpesh Ravindra Patil1, Takeshi Doi2, Ratnam Venkata Jayanthi2, Swadhin Behera2 (1.Young Research Fellow, APL, VAiG, JAMSTEC, 2.APL, VAiG, JAMSTEC)

Keywords:Convolutional neural networks (CNN), Deep-learning, Core monsoon zone, Summer monsoon rainfall prediction, Deficient and excess years, Long lead times

In this study we propose a deep learning scheme for the prediction of summer monsoon rainfall during the Jun-Sep period over the core monsoon zone (65-88oE; 18-25oN) of India, the region of large variability in rainfall in the season. The dynamical and statistical models show modest predictions skills over this region due to highly non-linear nature of interactions between atmosphere-ocean-land in the generation of the monsoon rainfall. It would be interesting to verify if non-linear techniques such as deep learning can have better skills in predicting the seasonal monsoon rainfall. The deep learning model is trained using oceanic (sea-surface temperature, and sub-surface temperature) and atmospheric (sea-level pressure) anomalies over 2- to 17-month lead time during the period 1871 to 1990 and the predicted rainfall anomalies during the period 1992-2021 are validated against the anomalies derived from the Indian Institute of Tropical Meteorology (IITM) spatially averaged over core monsoon zone. The anomaly correlation coefficient (ACC) skill score shows significant values across all lead times with substantially higher at 5-, 8- and 12- month lead times. Interestingly, few extreme cases of rainfall were accurately predicted over these lead times, including the rainfall deficient years of 2002, 2009, 2014, 2015, 2017 and 2018; excess years of 2013 and 2019. Overall, the predictive skill was noted higher for the deficient rainfall years than excess years. The results of the study will be useful to complement the efforts of the forecasting centers.