Rainfall Nowcasting based on Satellite Images using Convolutional Long-Short Term Memory

Rain that falls at a small and moderate rate is a blessing but incessant rainfall can bring many adverse effects such as loss of life, and destruction of property, crops and fields. One of the ways that can reduce the negative effects of natural disasters like this is to study trends and make predictions about what will happen. The predictive system, e.g. a nowcasting model can play an important role in dealing with rain issues, especially being able to provide early warning before bad weather occurs and this to some extent can help save lives and property. However, the determination of predictive models is a technically challenging task because rainfall is a non-linear phenomenon. In this research, a combination of a deep learning model called Convolutional Long-Short Term Memory (ConvLSTM) assisted by digital image processing is applied to real-time radar time series and satellite images. The goal is to predict the next event in a sequence of images with different timestamps i.e., 10-minutes, 30-minutes, and 60-minutes. Experimental results were evaluated with performance metrics using the Structural Similarity Index Measure (SSIM).