A  recent collaborative study between the Indian Institute of Science (IISc), Bengaluru, and the Indian Institute of Technology (IIT), Kharagpur, has employed machine-learning techniques to reveal new predictors for the Indian monsoon, making monsoon predictions more reliable.

The team consisted of Moumita Saha and Professor Pabitra Mitra from the Department of Computer Science, IIT–Kharagpur; and Professor Ravi S Nanjundiah from the Centre for Atmospheric and Oceanic Sciences, Bengaluru, as well as Divecha Centre for Climate Change, IISc. Using global climate data from 1948-2000 and machine learning algorithms, the team derived a set of reliable predictors for monsoon rainfall of the sub-continent.

Predicting monsoon in India has always been a challenge due to the influence of climate change and India’s humid tropical climate. Typically, in all monsoon forecast models, sea surface temperature (SST) and sea level pressure (SLP) are used as predictors because they are important regulators of monsoon.

Existing models consider predictors from certain influencing regions of Atlantic and Indian oceans. However, global climatic changes lead to fluctuations on multiple scales. This could alter the effect of existing predictors on monsoon phenomenon and new climatic relationships could evolve. Hence, it is necessary to continuously search for new predictors that influence the Indian monsoon and include them in building new deep-learning computational models. 

In the patterns

The researchers of the study tried to address this issue by collating the SST and the SLP values from all across the world from 1948 to 2000. They employed a deep- learning computational algorithm named ‘stacked autoencoder’ that could process such massive amount of data and identify global predictors for developing monsoon forecast models. A stacked autoencoder has multiple layers of calculations and every layer learns patterns in its input data by combining all the data points non-linearly. These patterns are learned and stored at the nodes of each internal layer. Nodes of the first layer serve as input to the second layer, and so the process continues. Greater the number of layers in the designed stack, greater is the complexity of the patterns revealed by the algorithm.

“We tried implementing a couple of neural networks,” reminisces Moumita, the lead researcher. “But the stacked autoencoder performed better in terms of accuracy and catching the extremes,” she explains. The team’s three-layer stacked autoencoder had multiple hidden nodes that represented complex relationships between global SSTs and SLPs.

Nodes that exhibited the highest correlation with Indian monsoon in the same period (1948-2000) were then chosen as the final set of predictors. Their study is the first report of a stacked autoencoder being used to identify predictors for the Indian monsoon. The researchers tested the newly identified predictors to match rainfall values for the period 2001-2014. Tests revealed a greater accuracy in prediction than existing models for both phases of the Indian monsoon — early (June-July) and late (August-September).

It was found that the new model forecasted early monsoon with mean absolute error of 6.8% in January, and the late monsoon with mean absolute error of 4.9% in March. Overall, their predicted rainfall values were closer to the real long period average (LPA) of rainfall as compared to Indian Meteorological Department’s (IMD) predictions.

Even for the present year, IMD predicted high probability of higher-than-normal rainfall. This new technique, on the other hand, predicted slightly below average rainfall. Ravi points out, “Current trends of monsoon behaviour indicate that predictions based on our technique could possibly be nearer to the actual value.” The team also designed separate stacked autoencoders that used global sea level pressures and sea surface temperatures independently. From the patterns obtained through these stacks, they derived distinct sets of monsoon predictions and found that the SLP-based model improved early monsoon prediction, whereas the SST-based model was most accurate in predicting the extreme cases — drought and excess rain.

This is a vastly different approach to that of the Indian Meteorological Department (IMD), which typically works with finite physics-based models as their meteorologists know the basic physical processes behind the monsoon. Hybrid models are the need of the hour. “We would like to collaborate with meteorologists and try to implement a hybridisation of the two approaches,” concludes Moumita on the next steps. Ravi proposes to extend this work to improve predictions on other scales as well. “This appears to be just the beginning. We expect the technique to give us many more interesting results,” he says.

(The author is with Gubbi Labs, a Bengaluru-based research collective)