YujiYamaguchi/lfm2-5-vl-450m-wildfire-precursor

FireGuard — Pre-Fire Vegetation Drought Detection

LFM 2.5-VL-450M LoRA fine-tuned on Sentinel-2 NDMI time-series imagery

Detects vegetation drought stress 7–21 days before wildfire ignition using three-point time-series of Sentinel-2 false-color composites.

Built for the Liquid AI × DPhi Space "AI in Space" Hackathon 2026.

Model Details

• Base model: LiquidAI/LFM2.5-VL-450M
• Fine-tuning method: LoRA (r=8, alpha=16, dropout=0.05)
• Target modules: q/k/v/out_proj (attention), w1/w2/w3 (FFN), in_proj, linear_1/2 (multimodal projector)
• Training: 5 epochs, lr=2e-4, effective batch=4, bfloat16, RTX 5090

What It Does

Given three Sentinel-2 false-color composite images of the same chaparral location at −21 days / −14 days / −7 days before a reference date, the model predicts:

{"risk_level": "HIGH"}   // vegetation progressively drying — elevated fire fuel risk
{"risk_level": "LOW"}    // stable or moist vegetation

Image encoding (R=B12/SWIR22, G=B8A/NIR, B=B11/SWIR16):

• Blue channel darkening toward −7d → NDMI declining → drying stress signal

Evaluation Results

Test set: n=12 held-out chaparral scenes (6 HIGH / 6 LOW), California, 2017–2023

Metric	Base LFM2.5-VL (zero-shot)	FireGuard LoRA
Precision	0.500	1.000
Recall	1.000	1.000
F1	0.667	1.000
FP Rate	1.000	0.000
Latency	283 ms	208 ms

The base model predicts every scene as HIGH risk (FP Rate = 1.0). Fine-tuning is essential.

Training Data

• Dataset: YujiYamaguchi/fireguard-sentinel2-wildfire-precursor
• POS: NASA FIRMS VIIRS_SNPP_SP California chaparral fire events → SimSat Sentinel-2 scenes at −21d/−14d/−7d before ignition
• NEG: Same coordinates, fire-free reference dates (FIRMS-verified)
• Split: train=100 / val=12 / test=12 (stratified, seed=42)
• Vegetation: California chaparral only (NDMI signal validated at p=0.038 for −7d lead time)

Usage

from transformers import AutoProcessor, AutoModelForImageTextToText
from peft import PeftModel

base = AutoModelForImageTextToText.from_pretrained("LiquidAI/LFM2.5-VL-450M")
model = PeftModel.from_pretrained(base, "YujiYamaguchi/lfm2-5-vl-450m-wildfire-precursor")
processor = AutoProcessor.from_pretrained("LiquidAI/LFM2.5-VL-450M")

See the FireGuard repository for full inference pipeline.

Citation

@misc{fireguard2026,
  title   = {FireGuard: Pre-Fire Vegetation Drought Detection via Sentinel-2 NDMI Time-Series},
  author  = {Yuji Yamaguchi},
  year    = {2026},
  note    = {Liquid AI × DPhi Space AI in Space Hackathon}
}