Enhancing Spatio-Temporal Forecasting with Spatial Neighbourhood Fusion:A Case Study on COVID-19 Mobility in Peru

TL;DR

A lightweight and model-agnostic Spatial Neighbourhood Fusion (SPN) technique that augments each cell's features with aggregated signals from its immediate H3 neighbors provides a simple yet effective path toward robust spatio-temporal forecasting during public health crises.

Abstract

Accurate modeling of human mobility is critical for understanding epidemic spread and deploying timely interventions. In this work, we leverage a large-scale spatio-temporal dataset collected from Peru's national Digital Contact Tracing (DCT) application during the COVID-19 pandemic to forecast mobility flows across urban regions. A key challenge lies in the spatial sparsity of hourly mobility counts across hexagonal grid cells, which limits the predictive power of conventional time series models. To address this, we propose a lightweight and model-agnostic Spatial Neighbourhood Fusion (SPN) technique that augments each cell's features with aggregated signals from its immediate H3 neighbors. We evaluate this strategy on three forecasting backbones: NLinear, PatchTST, and K-U-Net, under various historical input lengths. Experimental results show that SPN consistently improves forecasting performance, achieving up to 9.85 percent reduction in test MSE. Our findings demonstrate that spatial smoothing of sparse mobility signals provides a simple yet effective path toward robust spatio-temporal forecasting during public health crises.

Figures (2)

• Figure 1: Urban mask construction guided by night-time lights. Raw VIIRS radiance is first aligned with the H3 grid system, yielding a binary mask where cells above the radiance threshold (8nW cm^-2 sr^-1) are marked as urban. This binary urban mask is then smoothed using kernel density estimation to form contiguous urban clusters. The resulting H3-smoothed mask enables downstream models to differentiate between urban and peri-urban dynamics.
• Figure 2: End-to-end forecasting pipeline. An SPN block (left, not shown) fuses each H3 cell with the median of its six neighbours; the resulting three-channel tensor ($L\times3$) is patch-embedded and fed into one of three backbones. NLinear applies channel-wise linear layers; PatchTST operates on non-overlapping patches with a Transformer encoder; Kernel U-Net (centre) uses a symmetric encoder–decoder whose blocks wrap custom kernels (Linear, MLP, LSTM, Transformer). Skip connections preserve high-resolution context, and the latent vector is jointly down-sampled along time and channel dimensions. All backbones output the predicted mobility flow for the next $T$ hours.