Resilient Computer Vision for Blowing Snow

Reliable visual perception is a functional requirement for safe autonomous and human-driven transportation, yet in high-altitude regions like Wyoming, this capability is frequently compromised by blowing snow. While standard falling snow is a well-studied degradation in computer vision, blowing snow represents a novel and significantly more complex challenge. It involves high-velocity winds that suspend fine ice particles, creating dense volumetric scattering during the day and pronounced light bloom around artificial sources at night. In environments like Wyoming's I-80 corridor, these "whiteout" conditions create a severe safety risk that traditional de-noising models are unequipped to handle.

The core scientific hurdle is the non-stationary nature of the degradation. Unlike static rain or light snow, blowing snow dynamically obscures scene geometry in a way that traditional filters cannot easily resolve. Our research focuses on a framework designed to isolate the intrinsic scene information from these intense environmental perturbations. By treating the weather as a distinct, separable signal, the system can "see through" high-mass blizzard concentrations to recover underlying infrastructure. This ensures that Digital Twins maintain high-fidelity synchronization even during the extreme, unpredictable atmospheric events characteristic of the Mountain West.

Visualizing the transition from basic de-noising to robust structural restoration on extreme "whiteout" blizzard samples.

Universal Adverse-Weather Restoration

In traditional computer vision, "all-weather restoration" is a significant hurdle because different weather conditions, such as rain, fog, snow, or low-light, affect image pixels through vastly different physical processes. Rain adds directional streaks, fog causes uniform contrast loss, and low-light introduces non-stationary sensor noise. Most existing solutions require separate, heavy models for each condition, which creates a "computational bottleneck" for resource-constrained rural infrastructure.

For a Digital Twin to be effective, it requires a continuous, clean data stream from the physical world. If the sensors in Laramie or Afton are "blinded" by a sudden transition from fog to a night-time blizzard, the Digital Twin loses its synchronization with reality. To maintain a Trustworthy AI framework, we develop lean, unified models that can identify specific weather degradations and adapt restoration strategies in real-time without overwhelming the hardware.

Our approach addresses the efficiency gap in multi-weather restoration by utilizing multimodal embedding and hybrid attention backbones to handle diverse degradations within a single architecture. By optimizing the parameter space, we achieve a 56.5% reduction in model size compared to current benchmarks, making these solutions ideal for deployment on the low-power edge sensors used in our Laramie and Afton pilots.

CDD-11 Dataset: Input / Output / Ground Truth

LOL Dataset: Input / Output / Ground Truth

SNOW100K Dataset: Input / Output / Ground Truth

Predictive Digital Twins for Rural Mobility

Parking and transportation management in rural and medium-scale retail hubs face a critical bottleneck: a lack of proactive foresight. While real-time sensing provides status updates, most systems remain reactive, offering no path for predictive facility management. From a Digital Twin perspective, the primary challenge is decision-fragility in environments where historical data is scarce. Current frameworks often assume access to massive datasets or high-bandwidth cloud pipelines, conditions rarely met in localized rural infrastructures.

Without a synchronized virtual mirror that can simulate "what-if" scenarios, facility managers are unable to validate operational policies or forecast turnover in real-time. The research problem involves bridging this gap, transforming descriptive monitoring into a predictive intelligence layer. This requires creating a framework that can synthesize realistic behavior patterns and deliver multi-horizon forecasts within resource-constrained environments, ensuring the virtual environment remains a faithful and useful mirror of the physical asset's future state.