cmpatino/landmine-detection

Landmine Detection in Antioquia, Colombia

Dataset Description

This dataset supports landmine risk estimation across 15 municipalities in the Antioquia department of Colombia. The region is divided into 26,761 grid cells of 500 × 500 meters each, with 63 predictive features derived from multiple open geospatial, socio-economic, and conflict-related data sources. Each grid cell is labeled with a binary outcome indicating whether landmines were found during humanitarian demining operations.

The dataset is heavily imbalanced: only 1.6% of grid cells are positive (landmine present), reflecting the real-world rarity of contaminated areas even in conflict-affected regions.

Data from:

Dulce Rubio et al. "RELand: Risk Estimation of Landmines via Interpretable Invariant Risk Minimization." ACM Journal on Computing and Sustainable Societies 2.2 (2024): 1–29.

Usage

Loading with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("cmpatino/landmine-detection")
df = ds["train"].to_pandas()

# Features and target
X = df.drop(columns=["LATITUD_Y", "LONGITUD_X", "Municipio", "mines_outcome"])
y = df["mines_outcome"]

print(f"Samples: {len(df)}, Features: {X.shape[1]}, Positive rate: {y.mean():.3%}")
# Samples: 26761, Features: 63, Positive rate: 1.607%

Loading directly with pandas

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/cmpatino/landmine-detection/data/train-00000-of-00001.parquet"
)

# Quick overview
print(df.shape)            # (26761, 67)
print(df.dtypes.value_counts())
# float64    30
# int64      37

# Target distribution
print(df["mines_outcome"].value_counts())
# 0    26331
# 1      430

# Contamination rate per municipality
print(
    df.groupby("Municipio")["mines_outcome"]
    .mean()
    .sort_values(ascending=False)
    .head()
)

Dataset Summary

Property	Value
Grid cells	26,761
Features	63 continuous + categorical (one-hot encoded)
Target	mines_outcome (binary: 0 = no mines, 1 = mines found)
Positive rate	1.61% (430 positive cells)
Municipalities	15 (all declared mine-free after comprehensive demining)
Grid resolution	500 × 500 meters
Region	Antioquia, Colombia

Background

During the decades-long armed conflict in Colombia, anti-personnel landmines were extensively used by illegal armed groups to protect areas of strategic interest: coca crops, drug trafficking routes, laboratories, and illegal mining sites. Between 1990 and 2022, approximately 38,000 landmine events were recorded, resulting in over 12,000 victims (~20% fatal). Antioquia is the department with the highest historical number of landmine events and victims in Colombia.

The 15 municipalities in this dataset have been declared mine-free after comprehensive humanitarian demining operations, meaning the ground truth labels are based on actual demining results rather than estimates:

Abejorral, Alejandría, Chigorodó, Cocorná, El Carmen de Viboral, Granada, La Unión, Nariño, Sabanalarga, San Carlos, San Francisco, San Luis, San Rafael, San Roque, Sonsón.

Label Assignment

A grid cell receives a positive label (mines_outcome = 1) if it intersects at least one confirmed hazardous area or a cleared area where landmines were found during humanitarian demining operations, using data from Colombia's National Mine Action Center (AICMA). Otherwise, the cell receives a negative label (mines_outcome = 0), as these areas have been thoroughly inspected and declared free of mines.

Features

The 67 columns are organized as follows:

Identifiers & Target (4 columns)

Column	Description
LATITUD_Y	Latitude of grid cell center
LONGITUD_X	Longitude of grid cell center
Municipio	Municipality name (15 unique values)
mines_outcome	Target variable — 1 if landmines found, 0 otherwise

Continuous Features (30 columns)

These fall into three categories based on the RELand framework:

Remnants of War Indicators

Features capturing the historical presence of armed groups and conflict-related infrastructure.

Column	Description	Unit	Source
dist_old_mine	Distance to nearest historical landmine accident	km	DAICMA (2022)
coca_dist	Distance to nearest coca cultivation area	km	UNODC (2021)
hist_mines	Number of historical landmine accidents in municipality	incidents	DAICMA (2022)
No. Víctimas por Declaración	Declared conflict victims in municipality	people	Victims Unit (2019)

Geographic & Environmental Features

Features describing the terrain, climate, and ecological characteristics of each grid cell.

Column	Description	Unit	Source
elevation	Elevation	m above sea level	WorldClim (2023)
rainfall	Average annual rainfall	mm	WorldClim (2023)
temperature	Average temperature	°C	WorldClim (2023)
waterways_dist	Distance to nearest waterway	km	OpenStreetMap (2022)
soil_texture15_trans1	Soil texture component 1 (ILR-transformed)	—	Varón-Ramírez et al. (2022)
soil_texture15_trans2	Soil texture component 2 (ILR-transformed)	—	Varón-Ramírez et al. (2022)
altura	Average altitude of the municipality	m	Acevedo & Bornacelly (2014)

Socio-Demographic & Infrastructure Features

Column	Description	Unit	Source
population_2012	Population estimate for the grid cell (2012)	people	WorldPop (2023)
airports_dist	Distance to nearest airport	km	OpenStreetMap (2022)
seaport_dist	Distance to nearest seaport	km	OpenStreetMap (2022)
settlement_dist	Distance to nearest settlement	km	OpenStreetMap (2022)
finance_dist	Distance to nearest financial institution	km	OpenStreetMap (2022)
edu_dist	Distance to nearest school/educational institution	km	OpenStreetMap (2022)
buildings_dist	Distance to nearest building	km	OpenStreetMap (2022)
dist_roads_t1	Distance to nearest primary road	km	IGAC/Codazzi (2023)
dist_roads_t2	Distance to nearest secondary road	km	IGAC/Codazzi (2023)
dist_roads_t3	Distance to nearest tertiary road	km	IGAC/Codazzi (2023)
nighttime_lights_2012	Nighttime light intensity (2012)	index 0–63	Li et al. (2020)
retro_pobl_tot	Total municipal population	people	Acevedo & Bornacelly (2014)
indrural	Rural index of municipality	0 = urban, 1 = fully rural	Acevedo & Bornacelly (2014)
areaoficialkm2	Official municipal area	km²	Acevedo & Bornacelly (2014)
discapital	Distance to departmental capital (Medellín)	km	Acevedo & Bornacelly (2014)
pib_percapita_cons	GDP per capita (2009 constant prices)	millions COP	Acevedo & Bornacelly (2014)
rwi	Relative Wealth Index	index	Meta (2022)

Note: rwi (Relative Wealth Index) is listed in the paper's Appendix B feature table but is not present as a standalone column in this particular dataset version.

One-Hot Encoded Categorical Features (33 columns)

These are binary (0/1) indicator columns derived from three categorical variables:

Land Use (7 categories, source: IGAC 2022)

land_use_Agroforestal, land_use_Agrícola, land_use_Conservación de Suelos, land_use_Cuerpo de agua, land_use_Forestal, land_use_Ganadera, land_use_Zonas urbanas

Weather/Climate Zone (13 categories, source: IGAC 2022)

weather_Cuerpo de agua, weather_Cálido húmedo, weather_Cálido húmedo a muy húmedo, weather_Cálido muy húmedo, weather_Cálido seco, weather_Cálido seco a húmedo, weather_Frío húmedo, weather_Frío húmedo a muy húmedo, weather_Frío húmedo y frío muy húmedo, weather_Frío muy húmedo, weather_Muy frío y muy húmedo, weather_Templado húmedo a muy húmedo, weather_Zona urbana

Topographic Relief (13 categories, source: IGAC 2022)

relief_Abanicos aluviales subactuales y recientes, relief_Abanicos aluvio-torrenciales, abanicoterrazas y glacís subrecientes y antiguos, relief_Cuerpo de agua, relief_Espinazos, relief_Espinazos y colinas, relief_Filas y vigas, relief_Glacís coluvial y coluvios de remoción, relief_Glacís y coluvios de remoción, relief_Lomas y colinas, relief_Plano de inundación, relief_Plano de inundación y terrazas bajas, relief_Terrazas, relief_Terrazas y abanicos terrazas, relief_Vallecitos, relief_Vallecitos coluvio-aluviales, relief_Zona urbana

Municipality Distribution

Municipality	Grid Cells	% of Total
Sonsón	5,461	20.4%
Chigorodó	2,937	11.0%
San Carlos	2,911	10.9%
Abejorral	2,051	7.7%
San Luis	1,728	6.5%
San Roque	1,718	6.4%
El Carmen de Viboral	1,716	6.4%
San Rafael	1,478	5.5%
San Francisco	1,450	5.4%
Nariño	1,287	4.8%
Sabanalarga	1,074	4.0%
Cocorná	980	3.7%
Granada	773	2.9%
La Unión	678	2.5%
Alejandría	519	1.9%

Validation Protocols

The RELand paper proposes three validation protocols that account for the spatial structure of the data and simulate real demining operations:

1. blockCV — Block cross-validation within the Antioquia department, respecting municipal boundaries as spatial blocks.
2. blockV — Train on a subset of municipalities and validate on held-out municipalities (simulates deploying to a new area within the same department).
3. blockV-OOD — Train on Antioquia and evaluate on a completely different department (out-of-distribution generalization).

The Municipio column enables implementing these spatial validation schemes.

Ethical Considerations

This dataset is intended for research purposes in humanitarian demining and conflict-affected area analysis. The locations in this dataset correspond to municipalities that have already been declared mine-free. Predictions from models trained on this data should not be used as the sole basis for demining operations without expert human review and validation in the field.

Citation

@article{rubio2024reland,
  title={RELand: Risk Estimation of Landmines via Interpretable Invariant Risk Minimization},
  author={Rubio, Dulce and Patino, Camilo and Ramírez, Juan Felipe and Weidmann, Nils B and Müller, Emmanuel},
  journal={ACM Journal on Computing and Sustainable Societies},
  volume={2},
  number={2},
  pages={1--29},
  year={2024},
  publisher={ACM},
  doi={10.1145/3657055}
}

License

This dataset is released under the CC-BY-4.0 license.