A Framework for Multiscale Sensor Fusion in Environmental Monitoring

Thesis Abstract

This study addresses the challenge of accurately monitoring environmental conditions across heterogeneous landscapes by developing a framework for multiscale sensor fusion that integrates data from satellite, airborne, and ground-based sensors to improve situational awareness and decision support. The aim is to design a scalable sensor fusion framework that optimally combines multisource data at varying spatial, spectral, and temporal resolutions to generate high-fidelity environmental state estimates. Specific objectives include (1) to identify the theoretical underpinnings of multiscale data fusion in environmental monitoring and map them to a unified framework; (2) to formulate a data assimilation and fusion architecture that integrates proximal and remote sensing streams using Bayesian hierarchical modeling and probabilistic graph fusion; (3) to implement a modular prototype that demonstrates fusion performance across air quality, land-cover change, and hydrological indicators; (4) to evaluate the framework against ground truth and legacy datasets, and (5) to provide operational guidelines for deployment in real-world monitoring networks. A mixed-methods approach is employed. The research uses a quantitative design for the fusion framework development and validation, complemented by qualitative expert review for framework usability. The population comprises three sensor ecosystems satellite-based hyperspectral and multispectral systems (e.g., Sentinel-2, Landsat 8/9), airborne LiDAR and hyperspectral campaigns, and ground-based fixed and mobile sensor networks measuring air quality, soil moisture, and surface temperature. A stratified sampling strategy selects 15 study sites across urban, rural, and peri-urban environments to ensure diverse representativeness. Data collection instruments include satellite imagery (multispectral and hyperspectral data at 10–60 m resolution), airborne LiDAR point clouds, in-situ sensor arrays with calibrated PM2.5, PM10, NO2, O3, soil moisture, and temperature sensors, and meteorological station data. The data integration relies on a Bayesian hierarchical data assimilation framework augmented with a probabilistic graphical model to encode inter-sensor dependencies and temporal dynamics. Calibration and validation employ cross-comparison with high-accuracy ground truth and established datasets for air quality, reference-grade stations; for land cover, high-resolution field surveys; for hydrology, stream gauge records. Model specification includes a state-space representation for environmental indicators with latent variables representing true state, observation models for each sensor modality, and a fusion layer that optimally weights sources under uncertainty. Analytical techniques include Kalman filter variants (extended and unscented) for temporal fusion, particle filtering for non-linearities, and variational inference to manage computational complexity. Performance metrics encompass root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R2), as well as fusion-specific measures such as information gain and uncertainty reduction. Sensitivity analyses examine the impact of sensor density, data latency, and missing data on fusion accuracy. The study also employs thematic analysis of expert feedback to assess usability and interpretability of the framework. Key expected findings include demonstrable improvements in estimation accuracy for across-scale indicators (e.g., up to 25% RMSE reduction for PM2.5, 15–20% improvement in land-cover classification accuracy, and better groundwater-related indicators under sparse ground data conditions). The probabilistic fusion approach is anticipated to provide reliable uncertainty quantification and robust performance under sensor outages. The study contributes to knowledge by (a) articulating a formal multiscale sensor fusion architecture tailored to environmental monitoring, (b) advancing the integration of Bayesian data assimilation with graphical model fusion to handle heterogeneous data streams, and (c) delivering a validated prototype with transferable specifications for deployment in urban and regional monitoring networks. The framework also yields practical guidelines for sensor placement, data governance, and model maintenance to sustain performance over time. The main conclusion posits that a probabilistic, modular multiscale sensor fusion framework substantially enhances environmental state estimation and decision-support capabilities relative to single-sensor baselines, particularly in data-sparse or heterogeneous sensing environments. Recommendations include extending the framework to incorporate citizen science and participatory sensing for expanded spatial coverage, integrating real-time data streams for near-real-time monitoring, and pursuing open data standards to facilitate cross-agency interoperability.