AquaVeritas-LFM

AquaVeritas-LFM is a fine-tuned vision-language model for automated freshwater body monitoring from Sentinel-2 satellite imagery. It is a full fine-tune of LiquidAI/LFM2.5-VL-450M trained on 2,820 labeled observations across 20 global freshwater locations spanning 2018–2024.

Looking for GGUF? → Arty1001/aquaveritas-lfm-GGUF

Model Description

• Base model: LiquidAI/LFM2.5-VL-450M (Liquid Foundation Model 2.5 Vision-Language)
• Task: Satellite image analysis → structured JSON environmental assessment
• Fine-tuning: Full fine-tune (no LoRA/PEFT), 3 epochs on H100, 899 steps
• Training loss: 0.0113 | Eval loss: 0.01542
• Input: RGB + SWIR Sentinel-2 tiles (15 km × 15 km, 10 m/px)
• Output: Structured JSON with 10 environmental fields across core (water) and buffer (agriculture) zones

Output Schema

{
  "water_extent_status": "stable|shrinking|recovering|dry|flood_affected",
  "flood_risk": "none|low|moderate|high",
  "water_clarity": "clear|turbid|heavily_silted|algae_bloom",
  "shoreline_encroachment": true,
  "agriculture_present": true,
  "crop_stress_level": "none|low|moderate|severe",
  "crop_stress_type": "drought|waterlogging|none|mixed",
  "cultivation_expanding_toward_water": false,
  "settlement_visible": false,
  "bare_soil_expansion": false
}

Training Data

• 20 global freshwater locations across Africa, Asia, Europe, and the Americas
• 7-year temporal range: 2018–2024 (monthly Sentinel-2 observations)
• 2,820 training examples labeled by Claude Opus oracle (~99% field accuracy)

Evaluation

Field	Claude	Base LFM	AquaVeritas-LFM	Δ
Water Extent Status	86.7%	0.0%	100.0%	▲ 100%
Flood Risk	73.3%	33.3%	100.0%	▲ 67%
Water Clarity	93.3%	0.0%	100.0%	▲ 100%
Shoreline Encroachment	80.0%	50.0%	100.0%	▲ 50%