Sciamlab/landslide-prevention-italy-2024-blind-test

Pre-Landslide Risk Assessment 2024 Blind Test

Blind test dataset for evaluating VLM-based pre-landslide risk detection on 2024 Italian landslide events. This dataset contains 29 events from January–December 2024 with no ground truth labels — designed for evaluating model predictions against future landslide occurrences.

Generated for evaluating fine-tuned models on LiquidAI/LFM2.5-VL-450M.

---

Dataset Summary

Total samples	29
Split	test (blind)
Image resolution	512 × 512
Images per sample	6 (RGB + SWIR + NIR + NDMI + NDVI + BSI)
Geography	Italian regions, 2024 landslide events
Source data	Sentinel-2 L2A surface reflectance
Event period	January 1 – December 31, 2024

---

Dataset Structure

landslide_blind_test_2024/
├── README.md            ← this file
├── metadata.jsonl       ← one row per sample
└── images/              ← 6 input composites per sample
    ├── {id}_rgb.png
    ├── {id}_swir.png
    ├── {id}_nir.png
    ├── {id}_ndmi.png
    ├── {id}_ndvi.png
    └── {id}_bsi.png

{id} follows the pattern {event_id}_{town_slug}_{image_date} — e.g. EIT202400007_Genova_2023-12-22.

---

Data Fields

Field	Type	Description
id	string	Unique sample identifier (e.g., blind_EIT202400007)
location	string	Original event-folder location string
town	string	Town name
event_date	string (YYYY-MM-DD)	Date of the recorded landslide event
image_date	string (YYYY-MM-DD)	Sentinel-2 acquisition date
days_before_event	int	Gap between acquisition and event
split	string	test (blind test - no ground truth)
rgb_path … bsi_path	string	Relative paths to the 6 input images
output	string	Empty JSON (placeholder for model predictions)

---

Composite Images

Each sample contains 6 spectral composites generated from Sentinel-2 L2A:

Composite	Bands	Description
rgb	B04-B03-B02	Natural color
swir	B12-B8A-B04	SWIR composite (moisture/soil stress)
nir	B08-B03-B02	NIR false-color (vegetation health)
ndmi	(B08-B12)/(B08+B12)	Normalized Difference Moisture Index
ndvi	(B08-B04)/(B08+B04)	Normalized Difference Vegetation Index
bsi	((B11-B04)-(B08-B02))/((B11+B04)+(B08-B02))	Bare Soil Index

---

Usage

from datasets import load_dataset

# Load blind test dataset
ds = load_dataset("your-username/landslide-prevention-italy-2024-blind-test")

# Each sample has 6 images
sample = ds['test'][0]
print(sample['id'], sample['event_date'], sample['location'])

Evaluation

Use this dataset to evaluate pre-landslide detection models:

python scripts/evaluate.py \
  --hf-dataset your-username/landslide-prevention-italy-2024-blind-test \
  --model outputs/lfm2.5-vl-landslide-v2-Q8_0.gguf \
  --backend local

---

Source Data

• Imagery: Sentinel-2 L2A surface reflectance.
• Event ground-truth: Italian Emergency Events (EIT) database, 2024 landslide records.

---

Considerations for Using

• Blind test — ground truth labels are not provided. Model predictions should be compared against actual landslide occurrences after the fact.
• Spectral indices are proxies — NDMI shows surface moisture, not subsurface conditions.
• No DEM — slope steepness is inferred from visual cues.
• Temporal context — days_before_event provides the temporal gap between image acquisition and actual event.

---

Citation

If you use this dataset, please cite the Italian Emergency Events (EIT) database and the Copernicus Sentinel-2 mission.

---

Generated on 2026-05-02.