Can we use AI/ML to predict and understand the Indian Summer Monsoon
* Register (or log in) to the AI4G Neural Network to add this session to your agenda or watch the replay
Indian Summer Monsoon Rainfall (ISMR) has a huge impact on the life and economy of about 1.5 billion people living in India and its surroundings. Predicting ISMR at various scales is one of the toughest challenges in weather and climate. Traditionally the monsoon has been predicted using statistical models and more recently using dynamical models. We have attempted to use AI/ML for studying and predicting monsoons at various scales.
We will begin with an overview of the Indian Summer Monsoon and its variability at various scales and factors that affect it. On the seasonal scale, ISMR has a rich spatial structure influenced by warm seas around it and the Himalayan mountain ranges to its north. It is also influenced remotely by factors such as the El-Nino. Prediction of the seasonal mean monsoon with a long lead can help governmental agencies to take appropriate measures. On the intraseasonal scales, there are periods of intense rainfall called as ‘active spells’ interspersed with periods of low rainfall known as ‘break spells’. Prediction of these spells can be useful for the agricultural sectors and for hydrological operations (such for operating a reservoir or stream flow management). On still smaller spatial and temporal scales episodes of intense rainfall need to be predicted.
We will discuss prediction of ISMR on seasonal and interannual scales using AI/ML. We have used techniques such as stacked encoders for identification of potential predictors. We have used this in conjunction with regression tree models to predict the monsoon with the identified predictors. Our studies show that the skill is comparable to or better than dynamical models and is superior to conventional statistical models. We have also attempted to predict monsoon on shorter spatial scales such as a meteorological sub-division with skills superior to conventional techniques. Techniques such as LSTM and Seq2Seq methods have been used for predicting active and break spells. On daily scale we have used deep convolutional network (CNN) with UNET to predict rainfall. These and related topics will be discussed.