Domain-Aware Scalable Distributed Training for Geo-Spatiotemporal Data

Geo-spatiotemporal data has seen rapid growth due to advancements in data-generating techniques. In domains such as hydrology, ecology, and transportation, it is crucial to interpret complex spatiotemporal patterns using deep learning techniques. However, traditional deep learning models tend to provide suboptimal predictions without domain-specific knowledge and also require extensive computational resources to train on large-scale data. To address these challenges, we propose a novel approach for scaling deep learning for geo-spatiotemporal data using a domain-aware distributed training approach. Our network comprises a pixel-contribution block, a distributed multiheaded convolutional (CNN) spatial block, and a recurrent temporal block. We demonstrate the effectiveness of our approach using the case study of flood prediction in hydrology, where the network can effectively predict high peaks in discharge measurements at watershed outlets with up to 4. 1x speedup and up to 93% increase in prediction performance. Our empirical evaluation on a dataset of 23 watersheds in a northern region of the United States revealed an overall speedup of 22. 7x and an improvement of 16% in the mean prediction performance.