Uncontrolled wildfires are among the most destructive natural hazards, threatening lives, infrastructure, and ecosystems. A critical but understudied driver of wildfire spread is the ember, a burning particle that detaches from the main fire, travels with the wind, and ignites spot fires far ahead of the primary fire line. Because ember behavior is difficult to observe during real fires, it remains poorly understood and underrepresented in operational fire models.
This Smart and Connected Communities Integrative Research Grant (SCC-IRG) project develops a technological framework called Ember Intelligence, a new class of intelligent tools for detecting, tracking, and forecasting ember movement using drone-mounted infrared sensing, edge computing, and machine learning. The goal is to integrate fire science, unmanned aircraft systems, and distributed AI into a system that supports real-time situational awareness, improved fire prediction, and more effective decision-making for communities at risk.
The project is conducted in close collaboration with the US Department of Agriculture Forest Service and public safety stakeholders. Community engagement activities focus on regions near prescribed burn sites, including Richfield, Utah, to align technology development with operational needs and to strengthen community resilience. Educational outreach spans K–12 through postdoctoral levels, promoting participation in STEM and wildfire response research.
Machine learning models are becoming increasingly complex and computationally demanding, requiring significant resources for computation, storage, memory, and communication bandwidth. These requirements pose major challenges in tactical networks, where resources are constrained and network components are often mobile, distributed, intermittent, or disconnected.
Tactical AI applications are also increasingly multimodal. For example, an autonomous ground vehicle may perform image classification and audio analytics simultaneously for improved situational awareness. In addition, algorithms deployed in contested or hostile environments are vulnerable to adversarial attacks and system failures.
This project develops the Dynamic, Adaptive, and Swift AI at Tactical Networks (DAST) framework to address these challenges. DAST focuses on improving the speed, robustness, and security of distributed AI under severe resource constraints. The framework draws on tools from distributed computing, decentralized optimization, network protocols, coding theory, and information-theoretic security.
Learning in next generation wireless systems is expected to enable transformative applications across domains such as the Internet of Things, federated learning, mobile healthcare, and autonomous systems. In these settings, learning must be performed on data originating primarily at edge and user devices in order to reduce latency, improve resiliency, and enhance privacy.
Existing edge learning approaches typically rely on centralized cloud components to coordinate training and aggregate updates. This centralized structure can limit resiliency and create privacy vulnerabilities. This project advances a random walk based learning paradigm that relaxes strict centralization and enables a continuum between fully centralized and fully decentralized learning architectures.
In random walk learning, the model is treated as a baton that is updated and passed between nodes in the network according to adaptive policies. The project investigates theoretical and algorithmic challenges related to heterogeneity in data and network conditions, resiliency through redundancy and coding, model distribution across nodes via random walking snakes, and privacy preservation for locally owned data.