A Machine Learning Model for Active-Break Spell Forecasting of Indian Summer Monsoon from Outgoing Longwave Radiation Data

Forecasting the active/break spell of the Indian Summer Monsoon is important for planning flood management and irrigation activities. In the present work, we develop a simple machine learning model to forecast the active/break spells at multiple lead times. The input to the model is a sequence of 15 days of daily 1°× 1° gridded Outgoing Longwave Radiation (OLR) spatial maps. The spatial extents are from 30°S to 30°N and 80°W to 150°E. The output from the model is whether the next 5-day, 10-day and 15-day periods have active, neutral or break spells. The active/neutral spells are defined as a 5-day period with cumulative rainfall more/less than one standard deviation of the long term 5-day mean. This simple model consists of a linear dimensionality reduction in space and multiple classical classification models that use the principal components in time as input. This simple model achieves an F1 score of 0.46 for the first, second and third pentad forecasts. These results will serve as a common sense baseline for the development of any model based on complex neural networks. The code can be found at https://github.com/Shikamaru-Nara1729/OLR_PCA_Baseline.