Real-time prediction of spatial raster time series: a context-aware autonomous learning model

Real-time prediction of spatial raster time series, such as those derived from satellite remote sensing imagery, is important for making emergency decisions on various geo-spatial processes/events. However, because of the scalability issue and large training time requirement, the neural network (NN)-based models often fail to perform real-time prediction, in spite of their tremendous potential. In this paper, we propose ContRast, a variant of recurrent NN-based context-aware raster time series prediction model that attempts to resolve these issues by: (1) eliminating the need for offline adjustment of network structure by employing self-evolving autonomous learning of recurrent neural network, (2) saving training time by adopting single-pass parameter learning mechanism, and (3) reducing redundant learning by skipping sub-regional data associated with similar spatio-temporal context and reusing already learned parameters to predict for the same. Experimental evaluations with respect to predicting normalized difference vegetation index (NDVI)-raster derived from MODIS Terra satellite remote sensing imagery show that ContRast is highly effective for real-time prediction of spatial raster time series, and it significantly outperforms the existing models. In addition, the theoretical analyses of model complexity and computational cost further justify our empirical observations.