11:15 〜 11:30
[ACG34-09] Predicting East African short rains using convolutional neural networks
キーワード:East African Shot rains, Machine learning, Convolutional neural networks
East African region is known to experience severe floods/droughts during the season of short rains (Oct to Dec). The prediction of East African short rains (EASR) with sufficient lead-time will be quite helpful to plan mitigation efforts against its region-wide impacts. Here, we explored the seasonal predictability of average EASR during Oct to Dec (EASRI), using simple machine learning (artificial neural networks, ANN) and deep learning (convolutional neural networks, CNN) techniques. We trained the ANN’s from the past observed dipole mode index (DMI) values and CNN’s from the past global sea surface temperature (SSTA), vertically averaged (0-300m) sub-surface temperature (VATA) anomaly fields. These datasets for model training were obtained from centennial in situ observation-based estimates (COBE), Simple Ocean Data Assimilation (SODA) and Global Precipitation Climatology Centre (GPCC) for 82 years period (1901-1982) and tested on OISSTv2, Global Ocean Data Assimilation system (GODAS) and Global Precipitation Climatology Project (GPCP) for 38 years period (1983-2020).
The predictability of EASRI from these trained models was evaluated for various lead times, from 2 months (Sep initialized) to 6 months (May initialized). The ensemble mean prediction skills of EASRI from ANN’s and CNN’s were observed to be similar for 2 months lead-time; the correlation skills were 0.70 and 0.62, respectively. Whereas, a significant improvement in the skill at 6 months lead-time was found for CNN’s (correlation skill: 0.56) relative to ANN’s (correlation skill: 0.16). Furthermore, some extreme cases were successfully predicted such as catastrophic floods of 1997 and disastrous droughts of 2010 and 2016 at 6 months lead time. Thus, the results from this study emphasizes that an accurate long-lead prediction of East African short rains are possible using deep learning techniques.
The predictability of EASRI from these trained models was evaluated for various lead times, from 2 months (Sep initialized) to 6 months (May initialized). The ensemble mean prediction skills of EASRI from ANN’s and CNN’s were observed to be similar for 2 months lead-time; the correlation skills were 0.70 and 0.62, respectively. Whereas, a significant improvement in the skill at 6 months lead-time was found for CNN’s (correlation skill: 0.56) relative to ANN’s (correlation skill: 0.16). Furthermore, some extreme cases were successfully predicted such as catastrophic floods of 1997 and disastrous droughts of 2010 and 2016 at 6 months lead time. Thus, the results from this study emphasizes that an accurate long-lead prediction of East African short rains are possible using deep learning techniques.