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

FireEdge LoRA — Wildfire Detection from Sentinel-2 Imagery

LoRA fine-tune of LiquidAI/LFM2.5-VL-450M for binary wildfire detection from Sentinel-2 pseudo-color composites.

Developed for the Liquid AI × DPhi Space "AI in Space" Hackathon (2026).
Training code & full pipeline: github.com/YujiYamaguchi/liquid-ai-space-hackathon (apps/fireedge/)

---

Task

Given a Sentinel-2 pseudo-color image (R=SWIR22/B12, G=NIR/B8A, B=SWIR16/B11) of a 5km × 5km scene, predict whether an active wildfire is occurring.

{"fire_detected": true}   // or false

---

Model Details

Item	Value
Base model	LiquidAI/LFM2.5-VL-450M (450M params)
Adapter type	LoRA
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	q_proj, k_proj, v_proj, out_proj, in_proj, w1, w2, w3, linear_1, linear_2
Loss masking	Assistant-only (mask_asst)
Epochs	5
Learning rate	2e-4 (cosine, warmup_ratio=0.05)
Batch size (effective)	8 (per_device=1 × grad_accum=8)
Hardware	NVIDIA RTX 5090 (24GB VRAM)

---

Dataset

Split	POS (fire)	NEG (no-fire)	Total
Train	70	140	210
Val	15	30	45
Test	15	30	45

• POS: Active wildfire scenes sourced from NASA FIRMS VIIRS detections × Sentinel-2 via SimSat API
• NEG (FIRMS): Confirmed fire-free scenes near FIRMS detections (temporal offset)
• NEG (diverse): Geographically diverse non-fire scenes across multiple biomes
• GT source: NASA FIRMS VIIRS NRT — real satellite fire detections, no synthetic data
• Image format: 3-channel PNG (R=B12/SWIR22, G=B8A/NIR, B=B11/SWIR16), 224×224px

---

Evaluation Results

Tested on a held-out test set (45 samples, stratified split).

Metric	Base LFM2.5-VL	FireEdge LoRA	Δ
Precision	0.333	1.000	+0.667
Recall	1.000	0.933	−0.067
F1	0.500	0.966	+0.466
FP Rate	1.000	0.000	−1.000
Accuracy	0.333	0.978	+0.644
Latency mean (ms)	163	150	−13
JSON parse rate	1.000	1.000	—

Confusion matrix (FireEdge LoRA):

              Predicted
              NO-FIRE   FIRE
True NO-FIRE    30        0    ← 0 false alarms
True FIRE        1       14   ← 1 miss

The base model always predicts FIRE (FP Rate = 1.0). LoRA fine-tuning reduces FP Rate to 0.0 while maintaining Recall ≥ 0.85.

---

Training Methodology

Loss Masking Strategy

We compared two SFT loss strategies:

Strategy	Description	Recall	FP Rate
mask_asst (this model)	Loss only on assistant tokens	0.933	0.000
full_seq	Loss on all tokens	0.000	0.000

full_seq collapsed to always predicting false because the fixed system/user template tokens (242 tokens) dominate the sequence, leaving the 10-token assistant response (2.2% of total) with negligible gradient signal.

Input Format

<|im_start|>system
You are a satellite imagery analyst...
<|im_end|>
<|im_start|>user
[196 image tokens] Analyze this Sentinel-2 image...
<|im_end|>
<|im_start|>assistant
{"fire_detected": true/false}
<|im_end|>

---

Usage

from peft import PeftModel
from transformers import AutoProcessor
from PIL import Image

# Load base model + adapter
base_model_id = "LiquidAI/LFM2.5-VL-450M"
adapter_id = "YujiYamaguchi/lfm2-5-vl-450m-wildfire"

# See training code for full inference pipeline:
# apps/fireedge/fireedge/inference.py

Full inference code in the training repository.

---

Citation

@misc{yamaguchi2026fireedge,
  title  = {FireEdge: Onboard Wildfire Detection via LoRA Fine-tuning of LFM2.5-VL},
  author = {Yuji Yamaguchi},
  year   = {2026},
  note   = {Liquid AI × DPhi Space AI in Space Hackathon}
}