Enhancing Kriging with Inductive Spatio-Temporal GraphODE
Amin Sheykhzadeh, Behzad Moshiri, Ebrahim Ghafar‐Zadeh · 2024
Sensory networks in environmental monitoring provide real-time data on critical parameters, but the costs of installation and maintenance limit high-resolution data acquisition. Researchers aim to estimate values at specific locations without prior data samples, considering two approaches: virtual sensors and kriging. While virtual sensors face challenges in dynamic sensor networks where for every sensor added or disconnected the whole network should be retrained, kriging, especially spatio-temporal kriging using Graph Neural Networks, overcomes traditional kriging drawbacks and allows adaptability in dynamic sensor networks without frequent retraining. Despite their success, existing spatio-temporal kriging methods face challenges, notably the over-smoothing problem, restricting their ability to utilize deeper graph structures for a more comprehensive latent representation. In this paper, we propose a two-part method based on neural differential equations. The first part estimates values using spatial adjacency, while the second part refines these estimates considering temporal dependencies. Our approach explicitly addresses the over-smoothing problem, leading to a 2-8% improvement over state-of-the-art baseline methods. The results hold promise for enhancing the accuracy and effectiveness of environmental monitoring applications.