A multivariate spatiotemporal spread model of COVID-19 using ensemble of ConvLSTM networks

swarna kamal paul; Saikat Jana; Parama Bhaumik

A spatiotemporal epidemic spread model can accommodate both spatial and temporal correlations in data. However, most of the models either require disease specific domain knowledge [11] or are too spatially coarse [18] . Deep learning models can learn the dynamics of epidemic spread with high spatial resolution and high degree of accuracy with minimal initial bias due to its capability of highly nonlinear representation. Deep neural network based spatiotemporal models [4] have already been applied to predict epidemic spread. However, this model is experimented on a small localized region and influence of external factors are ignored. Deep learning models also tend to overfit to noise due to its high representational capability. Thus, modelling an epidemic spread in a wide region with high spatial and temporal resolution is challenging. To address the problem of spatiotemporal prediction of Covid-19 spread in a large geographical region with high resolution, we propose an ensemble of Convolutional LSTM [5] based model to be trained with multilayer temporal geospatial data, transformed as sequence of images. Each layer of the geospatial data corresponds to a causal factor that might influence the spread of the epidemic. We experimented with data of US and Italy and achieved country level mean absolute percent error (MAPE) of 5.57% and 0.3% respectively on forecasting of total infection cases in 5 days period. The paper is organized as following. In section 2 we conducted a literature review. A brief discussion on modelling the epidemic spread and data preparation method is presented in section 3. In section 4 we explained the ensemble of Convolutional LSTM model and performance measurement metrics. Section 5 is about experimental results. The following section concludes the paper.