Adaptive Spatial Prediction and Sensor Network Optimization for Environmental Monitoring: An Operations Research Approach

Abstract

Based on Information and Communication Technologies (ICT), predicting spatial data on irregular or global domains is challenging due to non-stationarity, asymmetric residual structure, non-Euclidean geometry, and sparse observations. In this research, an adaptive hybrid method is proposed to fuse the spherical Intrinsic Random Function (IRF) Kriging, skew-normal residual modeling, graph-Laplacian D-optimal sampling, adaptive model discrimination, and boldness-recalibrated probability prediction. The framework consists of removing low-frequency spherical-harmonic terms due to non-homogeneous structure, estimating an intrinsic covariance function for the truncated process, and applying constrained universal IRF Kriging. The predictive errors that are asymmetric are then modeled by a skew-normal residual layer. Both the graph-Laplacian regularization and the AIC/BIC-based discrimination are included in the sampling-design criterion to ensure informative spatial connectivity with sparse observations and to choose between different structural specifications. Finally, exceedance probabilities are transformed to the boldness scale to recalibrate them. The framework proposed here achieves lower point and probabilistic errors in the reported global temperature and irregular epidemiological experiments in cities, and significantly improves the efficiency of reduced sensor networks. Analysis of the results also indicates that the IRF order has the greatest influence, while the values of the parameters of graph regularization and probability recalibration are relatively stable across the tested range. The resulting architecture offers an integrated solution for spatial prediction in environments where one needs to strike a balance between accuracy, uncertainty quality, sampling efficiency, and decision usefulness.