Rainfall is an important meteorological phenomenon with economic and social consequences. Prediction of rainfall has thus gained much research attention. Rainfall prediction is done in different grains such as daily, monthly, etc. Since rainfall data is essentially time series data, methods like ARIMA and its variants have been extensively used for prediction. Recent research, however, focuses on deep learning networks for making predictions specifically using hybrid models exploiting both spatial and temporal information of data. Based on this idea, we decompose the rainfall into seasonality, trend, and noise components and exploit temporal features from seasonality and trend using LSTM and spatial features from noise using convolution. Further, we also add residuals from the GARCH model to improve prediction accuracy and process it using LSTM. Using rainfall datasets from 1901 to 2017 from three distinct geographical regions of India, we compare the prediction accuracy of our model with the standard ARIMA Model and find that our model provides better results. To further validate our model, we need to test the model for other regions and also use different data split methods.