MANGO: A Global Single-Date Paired Dataset for Mangrove Segmentation

This repository contains the official code for MANGO, a global single-date paired dataset for mangrove segmentation. This repository not only provides the MANGO dataset but also offers a comprehensive benchmark environment for training and evaluating various segmentation models on mangrove detection tasks.

Junhyuk Heo | Beomkyu Choi | Hyunjin Shin | Darongsae Kwon

Dataset

Data Collection Pipeline

The MANGO dataset is constructed using a quality-aware scene selection pipeline. For each site, multiple Sentinel-2 candidates are evaluated using the Fisher discriminant ratio to select the optimal single-date image that maximizes class separability between mangrove and background regions.

Dataset Overview

The MANGO dataset provides global coverage with country-disjoint train/validation/test splits for rigorous generalization testing. The dataset includes stratified sampling across different mangrove density levels.

Dataset Statistics

Split	Images	Masks
Train	34,272	34,272
Validation	4,159	4,159
Test	4,272	4,272
Total	42,703	42,703

• Image size: 256 x 256 pixels
• Image format: GeoTIFF (13 bands, uint16)
• Mask format: GeoTIFF (1 band, uint8, binary)

Sentinel-2 Band Information

The dataset contains 13-band Sentinel-2 imagery with the following bands:

Band Index	Band Name	Description	Resolution	Wavelength
1	B1	Aerosol	60m	443nm
2	B2	Blue	10m	490nm
3	B3	Green	10m	560nm
4	B4	Red	10m	665nm
5	B5	Red Edge 1	20m	705nm
6	B6	Red Edge 2	20m	740nm
7	B7	Red Edge 3	20m	783nm
8	B8	NIR	10m	842nm
9	B8A	Narrow NIR	20m	865nm
10	B9	Water Vapor	60m	945nm
11	B11	SWIR 1	20m	1610nm
12	B12	SWIR 2	20m	2190nm
13	SCL	Scene Classification	20m	-

RGB Visualization: To visualize images as true-color RGB, use bands 4 (Red), 3 (Green), 2 (Blue).

Benchmark Results

Quantitative Results

Benchmark results of segmentation models on the MANGO country-disjoint test set, comparing MVI-based and MF-based selection protocols.

Qualitative Results

Visual comparison of segmentation predictions across different baseline models.

Download

The MANGO dataset is available on Hugging Face:

https://huggingface.co/datasets/hjh1037/MANGO

Manual Download Instructions

1. Navigate to the Files and versions tab on the dataset page.
2. Find the data.tar file in the list.
3. Click the Download file icon (↓) on the right side of the file size.

Installation

After downloading, please move the file to your workspace and extract it. We recommend placing the dataset in the datasets/GEE/ directory to match the default configuration.

# Example: Create directory and extract
mkdir -p datasets/GEE
mv data.tar datasets/GEE/
cd datasets/GEE
tar -xvf data.tar