Multi-Sensor Data and Image Fusion for Environmental Parameter Prediction with Uncertainty-Aware Deep Learning

Abstract

Accurate environmental parameter prediction is essential for air quality management, precision agriculture, water resource protection, and urban climate adaptation. Traditional approaches rely on single-modality sensor data and point estimation, neglecting the complementary information available from remote sensing imagery and failing to quantify prediction uncertainty. This paper proposes MSAF-UQ, a Multi-Sensor Adaptive Fusion framework with Uncertainty Quantification for environmental parameter prediction. Three key innovations are introduced: (1) a Multi-Sensor Adaptive Fusion mechanism (MSAF) that employs self-attention alignment, cross-attention interaction, and adaptive gating to dynamically fuse heterogeneous sensor tabular data with satellite or UAV imagery, overcoming fixed-weight concatenation bottlenecks; (2) an Environment-Aware Uncertainty Decomposition (EAUD) framework that unifies Monte Carlo Dropout, deep ensemble learning, and conformal prediction to decompose total predictive uncertainty into epistemic, aleatoric, and distributional components, with a theoretically derived joint coverage probability lower bound; and (3) a Sensor-Weighted Adaptive Loss (SWAL) function that dynamically adjusts gradient contributions based on sample-level uncertainty and sensor reliability, enhancing learning on high-uncertainty samples. Systematic experiments on four cross-domain environmental benchmarks (Air Quality, Soil Moisture, Water Quality, and Urban Microclimate) demonstrate that MSAF-UQ achieves 23.1% and 25.8% improvements in RMSE and MAE, respectively, attains a PICP of 96.2% with 23.5% MPIW reduction, and exhibits strong cross-domain transferability with statistical significance (Wilcoxon rank-sum test, p < 0.01).