Abstract
As environmental monitoring systems increasingly rely on multi-sensor data, quantifying predictive uncertainty at the feature level is crucial for building robust and trustworthy deep learning models. In this study, we propose a perturbation-based framework for feature-wise uncertainty estimation in sensor networks, using the Intel Lab dataset as a case study. By applying controlled perturbations to individual input features and conducting Monte Carlo Dropout inference, the framework systematically quantifies how predictive uncertainty—computed as the mean variance over stochastic forward passes—varies as a function of perturbed feature values at each sensor location. This perturbation based mean uncertainty identifies critical value ranges where the sensor-specific model becomes less confident. Based on the continuously distributed spatial feature values obtained via inverse distance weighting interpolation, we further map the perturbation-based mean uncertainty onto the spatial domain to visualize the spatial distribution patterns of predictive uncertainty across sensor locations. Experimental results demonstrate that our method effectively reveals how a certain feature contribute to the mean model uncertainty offering actionable insights for data acquisition, sensor placement, and model refinement.