A deep learning approach for image time series forecasting: Study case, United States drought monitor

Image time series (ITS) are a chronologically ordered images set which contains spatial and temporal patterns. The United States drought monitor provides a map collection of drought severity spatial distribution across the regions which changes over the time. This work aims to use the drought map ITS to extract inner spatiotemporal features patterns and forecast the spatial drought severity distribution classes for a future horizon that ranges from one to twelve weekly time steps by using a convolutional long short-term memory network (ConvLSTM). This approach offers a new perspective by using a set of images (ITS) as input for a deep learning model to predict the spatial drought in an image that represents the next time step with the drought distribution. The design also allow us to implement a recursive multi-step forecasting strategy to generate an horizon up to twelve ($h$) weekly drought maps. The obtained results shows that the proposed approach achieves an overall of macro F1-score metric of 0.9953 in generating the next drought map $t+1$ and 0.6965 in generating the up to 12$th$ drought map. ConvLSTM is more accurate in general when it is compared with convolutional neural networks, video visual transformers networks and the naïve baseline model. These findings demonstrate the approach efficacy in identifying spatiotemporal features for reliable drought map forecasts, providing a new valuable tool for drought prediction as a visual representation.