26 - 30 April 2026
National Harbor, Maryland, US
Conference 14037 > Paper 14037-63
Paper 14037-63

Multispectral pre-fire hazard screening with an exponentially tilted L1 surrogate: loss-centered analysis on a fixed satellite scaffold

On demand | Presented live 28 April 2026

Abstract

This paper studies how a miss-sensitive objective alters the operating behavior of a compact satellite-sequence model for short-lead ignition screening on a fixed Indonesian FIRMS-Landsat scaffold. The analysis centers the exponentially tilted L1 family through gradient asymmetry, curvature, conditional risk, and threshold effects under severe class imbalance. Results show no universal best loss: BCE-style objectives lead on ranking and precision, focal loss yields the best F1, and tilted L1 is most useful in FN-averse alerting regimes where missed ignitions are costlier than false alarms. An α-sweep further shows a shift from all-negative behavior to recall-dominant screening, framing tilted L1 as a tunable decision loss rather than a general-purpose probability objective.

Presenter

Warren M. Xie
Singapore American School (Singapore)
Warren M. Xie is a student at Singapore American School who blends business, mathematics, and AI to build practical tools, including OctaHive, an AI-enabled smart beehive that streamlines monitoring and harvesting. He is also the founder and CEO of Nectarmas, a youth-led social enterprise that equips farmer cooperatives in Jambi with sustainable beekeeping to generate new income and reduce the risk of slash-and-burn fires. He collaborates with local farmer groups and private-sector partners to scale apiculture as a climate-positive livelihood model.
Application tracks: AI/ML
Presenter/Author
Warren M. Xie
Singapore American School (Singapore)