Liu et al. (2026) DeKNN: Decompositional Kriging Neural Network for efficient and interpretable spatiotemporal interpolation

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: GIScience & Remote Sensing
• Year: 2026
• Date: 2026-08-13
• Authors: Enbo Liu, Kaiqi Chen, Min Deng, Senzhang Wang
• DOI: 10.1080/15481603.2026.2705614

Research Groups

Not specified in the provided text.

Short Summary

The paper proposes the Decompositional Kriging Neural Network (DeKNN), a framework designed to efficiently and interpretably interpolate spatiotemporal data that is missing completely at specific locations (MCAL).

Objective

• To address the challenges of "missing completely at location" (MCAL) spatiotemporal interpolation by combining the nonlinear modeling capabilities of deep learning with the theoretical interpretability of geostatistical Kriging.

Study Configuration

• Spatial Scale: Not explicitly specified (general geographic information science applications).
• Temporal Scale: Not explicitly specified (time series analysis).

Methodology and Data

• Models used: Decompositional Kriging Neural Network (DeKNN), which comprises:
  • A position-aware gated network (Decomposer) to map observations to latent spatial features.
  • A latent-space Kriging neural network to estimate features at unsampled locations using distance-aware covariance structures.
  • A dual-branch reconstructor to recover complete time series via interpretable temporal basis vectors.
  • An adjoint Kriging mechanism for parameter learning.
• Data sources: Two real-world datasets.

Main Results

• DeKNN achieved superior interpolation accuracy compared to existing baseline models.
• The model substantially improved computational efficiency by shifting interpolation from high-dimensional spatiotemporal fields to a compact latent spatial space.
• The method provided interpretability through the analysis of learned latent spatial features and covariance structures.

Contributions

• Introduces a novel decomposition–interpolation–reconstruction architecture that mitigates the computational burden of high-dimensional spatiotemporal interpolation.
• Bridges the gap between "black-box" deep learning and interpretable geostatistics by integrating a Kriging system into a neural network framework.
• Specifically optimizes the handling of MCAL data, which is more challenging than "missing at random" (MAR) data.

Funding

Not specified in the provided text.

Citation

@article{Liu2026DeKNN,
  author = {Liu, Enbo and Chen, Kaiqi and Deng, Min and Wang, Senzhang},
  title = {DeKNN: Decompositional Kriging Neural Network for efficient and interpretable spatiotemporal interpolation},
  journal = {GIScience & Remote Sensing},
  year = {2026},
  doi = {10.1080/15481603.2026.2705614},
  url = {https://doi.org/10.1080/15481603.2026.2705614}
}