firesight-ir

FireSight-IR

A physics-informed neural network for wildfire detection and false-alarm rejection on satellite multispectral infrared data. Trained on VIIRS as a publicly available proxy for FireSat Protoflight's MWIR/LWIR instrument.

• Code: https://github.com/Ibekwemmanuel7/firesight-IR
• Author: Emmanuel Ibekwe (M.Sc. Atmospheric Science, Texas A&M University)

Model summary

Headline results (2023 validation)

Quick use

from huggingface_hub import hf_hub_download
import torch

ckpt = hf_hub_download(
    repo_id="emmanuelibekwe5525/firesight-ir",
    filename="firesight_pinn_best.pt",
)
# Then load with firesight_ir.FireSightPINN.from_checkpoint(ckpt)
# See https://github.com/Ibekwemmanuel7/firesight-IR for the package.

Ablation findings

Removing one branch or loss term per variant:

The surface branch (OSM proximity plus MODIS land cover) carries the discriminative signal. Physics loss contributes a regulariser.

Limitations

• Trained on western CONUS only; tropical and boreal regimes out of distribution.
• VIIRS-to-FireSat transfer requires validation on real FireSat data.
• The FA labelling rule (BTD < 20 K + OSM industrial proximity) and the physics thermal-realism loss (BTD < 10 K) are partially circular.
• No baseline comparison against threshold-only or VIIRS Collection 2.

Citation

@misc{ibekwe2026firesight,
  title  = {FireSight-IR: A physics-informed neural network pipeline for
            wildfire detection and false-alarm rejection},
  author = {Ibekwe, Emmanuel},
  year   = {2026},
  url    = {https://github.com/Ibekwemmanuel7/firesight-IR},
}