Autoencoder-based identification of predictors of Indian monsoon

Abstract

Prediction of Indian summer monsoon uses a number of climatic variables that are historically known to provide a high skill. However, relationships between predictors and predictand could be complex and also change with time. The present work attempts to use a machine learning technique to identify new predictors for forecasting the Indian monsoon. A neural network-based non-linear dimensionality reduction technique, namely, the sparse autoencoder is used for this purpose. It extracts a number of new predictors that have prediction skills higher than the existing ones. Two non-linear ensemble prediction models of regression tree and bagged decision tree are designed with identified monsoon predictors and are shown to be superior in terms of prediction accuracy. Proposed model shows mean absolute error of 4.5 % in predicting the Indian summer monsoon rainfall. Lastly, geographical distribution of the new monsoon predictors and their characteristics are discussed.