Synthesising datasets of varying uncertainty for spatial estimation

Abstract

Accurate spatial estimation is critical for decision-making in fields like agriculture, environmental science, and urban planning, yet integrating datasets with varying precision and uncertainty remains a challenge. This study evaluates three methodologies, Bayesian Hierarchical Models (BHM), Co-kriging, and Gaussian Process Interpolation (GPI), for synthesizing high-precision (e.g., ground sensors, LiDAR) and low-accuracy (e.g., satellite imagery, SRTM) spatial data. Using simulated and real-world datasets (crop yields, DEMs), we assess each method’s predictive accuracy, uncertainty quantification, computational efficiency, and practical applicability.

Results demonstrate that BHM excels in handling hierarchical data structures and uncertainty, though it is computationally intensive. Co-kriging performs well with correlated variables (e.g., rainfall and crop yield) but struggles with resolution mismatches (e.g., LiDAR vs. SRTM). GPI provides robust probabilistic predictions but faces scalability limitations. The study highlights trade-offs between accuracy and computational cost, recommending BHM for complex uncertainty-aware tasks and hybrid approaches for future work. These findings advance spatial data integration techniques, offering actionable insights for domain-specific applications.