VPint: value propagation-based spatial interpolation.

Area of Science:

• Geospatial data analysis
• Environmental modeling
• Data science

Background:

• Missing data is a prevalent issue in remote sensing, ecology, and meteorology.
• Existing spatial interpolation methods like Gaussian processes and spatial autoregressive models have limitations in capturing complex spatial interactions and assumptions about data homogeneity.

Purpose of the Study:

• To propose a new value propagation-based spatial interpolation method, VPint, to overcome the limitations of existing techniques.
• To introduce two variants of VPint: static discount (SD-MRP) and weight prediction (WP-MRP).

Main Methods:

• VPint utilizes a value propagation approach inspired by Markov reward processes (MRPs).
• The SD-MRP and WP-MRP variants operate locally while recursively incorporating global spatial relationships.
• Performance was evaluated against 8 common baselines using metrics like Mean Absolute Error (MAE) and Structural Similarity (SSIM) on synthetic and real-world datasets.

Main Results:

• VPint demonstrated a competitive advantage over baseline methods for randomly missing data, showing superior MAE and SSIM.
• For spatially clustered missing data, VPint achieved the best performance on 2 out of 3 datasets.
• Experiments confirmed the convergence and scalability of the proposed methods.

Conclusions:

• VPint offers an effective solution for spatial and spatio-temporal data interpolation, particularly in scenarios with missing data.
• The method's ability to implicitly account for global spatial relationships while operating locally provides a significant advantage.
• VPint shows promise for applications in various scientific fields requiring accurate spatial data imputation.