MoTiF: a self-supervised model for multi-source forecasting with application to tropical cyclones

Abstract We present a deep learning architecture that reconstructs a source of data at given spatio-temporal coordinates using other sources. The model can be applied to multiple sources in a broad sense: the number of sources may vary between samples, the sources can differ in dimensionality and sizes, and cover distinct geographical areas at irregular time intervals. The network takes as input a set of sources that each include values (e.g., the pixels for two-dimensional sources), spatio-temporal coordinates, and source characteristics. The model is based on the Vision Transformer, but separately embeds the values and coordinates and uses the embedded coordinates as relative positional embedding in the computation of the attention. To limit the cost of computing the attention between many sources, we employ a multi-source factorized attention mechanism, introducing an anchor-points-based cross-source attention block. We name the architecture MoTiF (multi-source transformer via factorized attention). We present a self-supervised setting to train the network, in which one source chosen randomly is masked and the model is tasked to reconstruct it from the other sources. We test