Edge-Inference Pipeline for Short-Horizon Wildfire Smoke Nowcasting

Abstract

This study evaluates a lightweight edge-inference workflow for near-real-time smoke-density nowcasting from low-power camera nodes. The authors compare quantized CNN variants under variable haze and lighting conditions and report latency, power draw, and forecast consistency across 600 annotated windows. The selected model reduced median inference latency by 31% while preserving prediction agreement within 4.2% of a larger cloud baseline. The paper outlines deployment trade-offs for constrained environmental sensing systems.