Mangrove mapping with SEPAL#

This article describes a complete workflow for mapping mangroves using SEPAL — FAO’s free, open-source, web-based cloud platform for satellite data analysis. The approach combines multi-sensor satellite data, global mangrove reference datasets, and machine learning classification to produce locally relevant mangrove maps and detect change over time. An optional extension covers above-ground biomass (AGB) mapping.

Note

Author: Erik Lindquist, Forestry Officer, Food and Agriculture Organization of the United Nations (FAO)

Overview#

The workflow consists of the following stages:

  1. Registration and setup

  2. Request SEPAL resources

  3. Define the area of interest

  4. Build an optical mosaic (optional)

  5. Build a radar mosaic (optional)

  6. Load global mangrove reference datasets

  7. Create the high-confidence mangrove reference map

  8. Generate training data

  9. Incorporate Google Satellite Embeddings

  10. Classify mangroves

  11. Map above-ground biomass (optional)

  12. Detect change using CCDC

  13. Download and export results

Why use SEPAL for mangrove mapping?#

SEPAL democratises access to satellite data and analysis tools, enabling autonomous land monitoring without the need for local compute infrastructure. For mangrove mapping specifically, SEPAL provides:

  • Multi-sensor data access – seamless integration of optical (Sentinel-2, Landsat, NICFI/Planet) and synthetic aperture radar (SAR) (Sentinel-1 C-band, ALOS L-band) imagery

  • Multi-temporal analysis – time-series processing and change detection using the Continuous Change Detection and Classification (CCDC) algorithm

  • Cloud-based classification – machine learning classification recipes powered by Google Earth Engine (GEE)

  • Global reference data – integration with existing global mangrove products to generate locally relevant training data automatically

  • Team collaboration – data saved as GEE assets can be shared across a team, allowing one person to build a composite that all team members can access and classify

1. Registration and setup#

Before starting, create accounts on two platforms.

1.1 Google Earth Engine#

SEPAL’s processing recipes run on GEE. A free GEE account is required.

  1. Go to earthengine.google.com.

  2. Sign up using a Gmail account.

  3. Wait for approval (usually within a few hours to a few days).

1.2 SEPAL#

  1. Go to sepal.io and register for a free account.

  2. Connect the GEE account within SEPAL settings.

See also

For a step-by-step walkthrough of SEPAL setup, see Register.

2. Request SEPAL resources#

SEPAL is always free to use; however, downloading data and running high-performance computing in the cloud requires resources to be allocated to an account. Resources are requested via the User report button (displayed as $ 0/h) in the lower-right corner of the SEPAL interface.

To request resources:

  1. Select the User report button in the lower-right corner of the SEPAL interface.

  2. Select Request additional resources.

  3. Enter the requested amounts. For most mapping work, the following is a reasonable starting point:

    • Processing: $10 USD/month of cloud compute time

    • Storage: $10 USD/month

    • Workspace: 10 gigabytes (GB) of personal SEPAL workspace storage

  4. Enter a short message describing the work (e.g. “Mangrove mapping”).

  5. Select Apply. The SEPAL team will review and approve the request.

Once resources are approved, the Download button (displayed as a cloud with an arrow icon) will become active in all recipes. Even $1 of resources is sufficient to activate downloading to GEE or Google Drive.

Note

SEPAL supports requests for larger amounts of resources for intensive work, such as training machine learning models. Describe the use case in the request message if needed.

3. Define the area of interest#

Use the same area of interest (AOI) for all downstream recipes to ensure composites, reference layers, and classification outputs are spatially aligned.

  1. Open Process in the SEPAL interface.

  2. Set the AOI using an administrative boundary, an uploaded geometry, or a drawn polygon.

4. Build an optical mosaic (optional)#

Note

This step is only required if Google Satellite Embeddings are not used as the classification input (see Section 9). If using embeddings, proceed directly to Section 6.

A cloud-free optical mosaic provides the spectral bands used for classification. In SEPAL, all processing starts with a recipe. Open a new recipe from the recipe book by selecting the + button, then the green Add recipe button.

Composite method#

Use the Medoid composite method, which selects actual observed pixel values rather than statistical summaries, preserving spectral integrity. Save the recipe with a meaningful name (e.g. Optical_Medoid_2023) so it can be referenced in the classification recipe.

See also

See Optical mosaics for full instructions on the SEPAL optical mosaic recipe.

5. Build a radar mosaic (optional)#

Note

This step is only required if Google Satellite Embeddings are not used as the classification input.

SAR data penetrates cloud cover and is sensitive to canopy structure and soil moisture — making it a valuable complement to optical data in tropical coastal environments where persistent cloud cover limits optical data availability.

6. Load global mangrove reference datasets#

Load three global datasets as EE Asset recipes in SEPAL to build the high-confidence reference map in Section 7. Each dataset is available in the GEE catalog or the Awesome GEE Community Catalog.

To add an EE Asset recipe:

  1. Add a new recipe and select EE Asset.

  2. Set the AOI.

  3. Search for the asset by name or paste the asset path.

  4. Select Done.

  5. To visualise, open the Menu (☰) → select + → select the band → set min/max values → select Apply.

  6. Save the recipe with the name indicated below.

6.1 Global Mangrove Watch (GMW) v3 — 2020 extent#

Asset path

projects/sat-io/open-datasets/GMW/extent/GMW_V3

Band

b1 (1 = mangrove, 0 = non-mangrove)

Date filter

2020-01-01 to 2020-12-31; select first image

Save as

GMW_2020_EEAsset

6.2 ESA WorldCover land cover — 2021#

Asset path

ESA/WorldCover/v200

Band

Map

Mangrove class value

95

Save as

WorldCover_2021_EEAsset

6.3 Digital elevation model (MERIT DEM)#

Asset path

MERIT/DEM/v1_0_3

Band

dem

Save as

MERIT_DEM_EEAsset

The digital elevation model (DEM) provides an elevation constraint — mangroves occupy low-lying intertidal zones, so pixels above 40 m elevation are excluded from the high-confidence product in Section 7.

7. Create the high-confidence mangrove reference map#

This step converts the three global datasets into a single high-confidence binary map of mangrove and non-mangrove, which is then used to automatically generate training data for the local classification.

If multiple independent global datasets agree that a pixel is mangrove (or non-mangrove), confidence in that label is much higher than if only one dataset indicates this. The elevation constraint further removes unlikely mangrove locations.

Remapping rules#

In SEPAL, use the Remapping recipe with the following logic:

Output class

Condition

Class 1 – Mangrove

GMW b1 = 1 AND WorldCover Map = 95 AND DEM dem < 40 m

Class 2 – Non-mangrove

GMW b1 = 0 AND WorldCover Map ≠ 95 AND DEM dem < 40 m

(No data)

Pixels that do not meet either condition are excluded

Output band: class (1 = mangrove, 2 = non-mangrove)

Save as: Mangrove_HighConfidence_Remapping

Note

The DEM < 40 m filter restricts both mangrove and non-mangrove samples to the intertidal zone, improving the relevance of the non-mangrove class and avoiding confusion with upland forest.

Tip

Use the divided interface (multi-view panel) to inspect the high-confidence product against the optical and radar mosaics before proceeding. Confirm that the mangrove extent looks reasonable for the area.

8. Generate training data#

With the high-confidence reference map produced, generate training data automatically by stratified sampling — more efficiently and at greater scale than manual digitising.

In SEPAL, training data is generated within the Classification recipe using the Sample classification option:

  1. Open a new Classification recipe.

  2. In the training data tab, select Add > Sample classification.

  3. Set the sampling source to Mangrove_HighConfidence_Remapping (recipe reference) or an exported GEE asset of the same (band: class).

  4. Use balanced (stratified) sampling — equal numbers of samples per class.

  5. Set a random seed for reproducibility.

9. Incorporate Google Satellite Embeddings (AlphaEarth Foundations)#

Note

These are referred to as “Alpha Earth Embeddings” in some training materials. They are now publicly available as the Google Satellite Embedding dataset, powered by AlphaEarth Foundations — a geospatial artificial intelligence (AI) model developed by Google and Google DeepMind.

What are satellite embeddings?#

Satellite embeddings are pre-computed, AI-generated feature vectors that summarise multi-sensor, multi-temporal Earth observation data into a compact representation. Each 10 m pixel is described by a 64-dimensional embedding vector encoding a full year of observations from:

  • Sentinel-2

  • Landsat

  • Sentinel-1 SAR

  • Copernicus DEM (elevation)

  • ERA5 (climate)

  • GEDI LiDAR (vegetation structure)

Key advantages for mangrove mapping:

  • Fewer training samples needed – embeddings encode rich spatial and temporal context, reducing the labelled points required for accurate classification;

  • Cloud-robust – embeddings summarise an entire year of acquisitions, making them more robust to cloud contamination than single-date composites;

  • No deep learning infrastructure required – pre-computed and analysis-ready; compatible with SEPAL’s built-in classifiers;

  • Faster classification – typically faster than classifying equivalent optical and radar inputs separately; and

  • Temporal coverage – annual embeddings available from 2017 onwards.

Loading embeddings as an EE Asset recipe#

Asset path

GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL

Asset type

ImageCollection — mosaic or select target year

Bands

All 64 embedding bands

Save as

AlphaEarth_Embeddings_EEAsset

  1. Add a new recipe and select EE Asset.

  2. Set the AOI.

  3. Search for embeddings and select Satellite Embeddings V1, or paste the asset path above.

  4. Select Done.

  5. To confirm data is loading: open the Menu (☰) → select + → select any band (e.g. Z00) → select Apply.

  6. Save the recipe as AlphaEarth_Embeddings_EEAsset.

When to use embeddings vs. optical/radar mosaics#

  • Use embeddings when an annual classification is needed and good results are required with limited training data or compute time. When embeddings are used, optical and radar mosaics are not needed as classifier inputs.

  • Use optical and radar mosaics when classification at a specific sub-annual time point is needed, or when within-year temporal detail is important.

  • Both approaches can be combined if desired.

See also

10. Classify mangroves#

With training data and image inputs prepared, run the SEPAL Classification recipe to produce a mangrove/non-mangrove map.

10.1 Set up the classification recipe#

  1. Add a new recipe and select Classification.

  2. When asked which image to classify, choose one of the following:

    • Embeddings approach: AlphaEarth_Embeddings_EEAsset only (do not add optical or radar mosaics)

    • Optical/radar approach: the optical mosaic recipe plus the radar mosaic recipe

    • Combined approach: embeddings plus optical and/or radar mosaics

10.2 Select input bands#

For the embeddings approach, add all 64 embedding bands.

For the optical/radar approach, recommended bands are:

  • Sentinel-2: B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12, plus derived indices (NDVI, NDMI, NDWI, Tasseled Cap)

  • Sentinel-1: VV mean, VH mean, VH standard deviation

10.3 Define classification classes#

For mangrove mapping, a three-class legend is recommended:

Class

Description

Mangrove

Mangrove forest

Non-mangrove

All other land cover

Water

Open water

Keep classes distinct — overlapping or ambiguous classes reduce classification accuracy. Assign recognisable colours (e.g. green for mangrove, blue for water).

10.4 Add training data#

Reference the Mangrove_HighConfidence_Remapping recipe (or exported asset) generated in Section 7 and sampled in Section 8. In the training data tab, select Add > Sample classification.

Add manual training points using the Marker icon if local knowledge suggests corrections are needed.

Tip

The classification updates in real time as training points are added or modified, allowing the impact of each change to be seen immediately.

10.5 Configure the classifier#

SEPAL uses Random Forest by default. Recommended settings:

Parameter

Recommended value

Number of trees

~300 (use 25 for exploration; increase for production)

Variables per split (mTry)

Default/auto

Probability outputs

Enable if available

10.6 Interpret and refine results#

Use the Menu (☰) to switch between display bands:

  • Class – the final classified map (one class per pixel)

  • Mangrove percent – the probability that each pixel belongs to the mangrove class; more informative than the binary class map for understanding model uncertainty

To improve results, iterate by:

  • tightening the high-confidence mask criteria in the Remapping recipe (Section 7) and resampling (Section 8);

  • adding more training samples;

  • increasing the number of Random Forest trees; or

  • adjusting the input data stack.

See also

See Classification for full instructions on the SEPAL classification recipe.

11. Map above-ground biomass (optional)#

In addition to mangrove extent, SEPAL can produce wall-to-wall maps of above-ground biomass (AGB) by combining point-based biomass measurements with satellite embeddings using a Regression recipe.

11.1 Approach#

Where biomass measurements are available (from field plots or GEDI LiDAR), and where the 64 embedding bands are available, the known biomass values can be statistically related to the embedding values and applied to every pixel across the AOI — producing a continuous biomass estimate map.

11.2 Option A – Using GEDI LiDAR data#

The Global Ecosystem Dynamics Investigation (GEDI) is a full-waveform LiDAR instrument on the International Space Station, capturing vegetation height, canopy structure, and above-ground biomass density. GEDI data are point samples rather than wall-to-wall, but can be used as training data to produce a continuous biomass map.

To load GEDI L4A as an EE Asset recipe:

Asset

GEDI L4A Above Ground Biomass Density v2.1

Band

AGBD (above-ground biomass density, Mg/ha)

Visualisation

Thermal colour palette, min = 0, max = ~194

Save as

GEDI_L4A_EEAsset

To run the biomass regression:

  1. Add a new recipe and select Regression.

  2. Select AlphaEarth_Embeddings_EEAsset as the image (all 64 bands).

  3. In the training data tab, select Add > Sample image.

  4. Select GEDI_L4A_EEAsset as the source, with AGBD as the target band.

  5. Set the number of samples (increase for production results).

  6. Select Apply and run.

Note

If the study area is changing rapidly, match the GEDI acquisition year to the target image year. For more stable mangrove areas, using all available GEDI data is acceptable.

11.3 Option B – Using field-collected AGB data#

Field-collected plot data with geographic coordinates can be uploaded to GEE as an asset table and used as training data in the Regression recipe.

To upload field data to GEE:

  1. In the GEE Code Editor, go to the Assets tab.

  2. Select New > CSV file.

  3. Upload the data file — GEE handles reprojection automatically.

  4. Note the asset ID from the asset details panel.

To run the regression using field data:

  1. Add a new recipe and select Regression.

  2. Select AlphaEarth_Embeddings_EEAsset as the image (all 64 bands).

  3. In the training data tab, select Add > Earth Engine table.

  4. Paste the field data table asset ID.

  5. Select the AGB column as the target value band.

  6. Select Apply and run.

11.4 Interpreting biomass results#

The regression produces a continuous map of estimated above-ground biomass (Mg/ha). Results improve significantly with more training samples and more trees in the Random Forest model. Consider reserving 20–30% of samples for independent validation.

12. Detect change using CCDC#

To map mangrove loss or gain over time, this workflow uses the Continuous Change Detection and Classification (CCDC) algorithm (Zhu and Woodcock, 2014). CCDC fits a time-series model to each pixel using all available satellite observations and detects structural breaks — making it well suited to detecting mangrove conversion events.

12.1 Create a CCDC asset#

  1. Add a new recipe and select CCDC Asset.

  2. Define the AOI and date range.

  3. Select input imagery (Sentinel-2 and/or Landsat recommended).

  4. Run the recipe and save the result as a GEE asset.

Tip

Pre-built CCDC assets for specific study areas may be shared by the SEPAL team or colleagues as GEE asset links, avoiding the need to run this computationally intensive step independently.

See also

See CCDC slice for full instructions on creating a CCDC asset.

12.2 Extract CCDC slices#

A CCDC slice is a snapshot of the fitted time-series model at a specific date — providing harmonised, cloud-free spectral values consistent across dates. CCDC slices can be used:

  • as additional input bands in the Classification recipe (Section 10) for improved accuracy using spatio-temporal descriptors; or

  • from two different dates to detect change between them.

See also

See CCDC slice for full instructions on using CCDC slices.

12.3 Run change detection#

Option A – Index Change recipe (two-date comparison):

  1. Add a new recipe and select Index Change.

  2. Load CCDC slices from two target dates (e.g. 2015 and 2023).

  3. Run the recipe to produce a map of positive and negative spectral change.

  4. Interpret positive and negative values as mangrove gain or loss.

Option B – Near-real-time alert recipe:

SEPAL includes an alert recipe based on CCDC that compares the time-series model to very recent observations (last few days to weeks), enabling near-real-time detection of mangrove disturbance events for operational monitoring.

13. Download and export results#

13.1 Export destinations#

Select the Download button (cloud with arrow icon) in any recipe to open the export dialog. Choose:

  • Bands to export – e.g. class for a classified map, or mangrove_percent for a probability map;

  • Scale – spatial resolution in metres (e.g. 10 m for Sentinel-2 or embeddings-based results); and

  • Destination:

    • Google Earth Engine asset – exports to the GEE asset folder; recommended for further analysis or sharing;

    • Google Drive – exports to Google Drive for local download; or

    • SEPAL workspace – exports to the SEPAL workspace; use if running further analysis inside SEPAL.

13.2 Monitor export tasks#

After selecting Retrieve, a spinning icon appears on the purple task button. A green check mark indicates completion. Monitor tasks in the GEE Code Editor under the Tasks tab.

13.3 Upload data to GEE#

To use local raster or vector data in SEPAL (e.g. field plots, administrative boundaries, existing maps):

  1. In the GEE Code Editor, go to the Assets tab and select New.

  2. Select the file type:

    • GeoTIFF for raster data;

    • Shapefile – upload .shp, .shx, .dbf, and .prj files together; or

    • CSV for tabular data with geographic coordinates.

  3. GEE handles reprojection automatically.

  4. Copy the asset ID and use it in any SEPAL EE Asset recipe or training data input.

Additional resources#

Resource

Link

SEPAL platform

sepal.io

SEPAL documentation

docs.sepal.io

SEPAL certified online course

FAO SEPAL course

FAO SEPAL brochure

fao.org/in-action/sepal

Global Mangrove Watch v3.0

JAXA GMW

ESA WorldCover

ESA WorldCover

GEDI L4A AGB dataset

GEDI L4A on GEE

GEE Satellite Embeddings catalog

GEE Data Catalog

AlphaEarth Foundations

Google DeepMind blog

Contact

sepal@fao.org

References#

Brown, C.F., Kazmierski, M.R., Pasquarella, V.J., et al. 2025. AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. arXiv preprint arXiv:2507.22291.

Simard, M., Fatoyinbo, L., Smetanka, C., Rivera-Monroy, V.H., Castañeda-Moya, E., Thomas, N. and Van der Stocken, T. 2019. Mangrove canopy height globally related to precipitation, temperature and cyclone frequency. Nature Geoscience, 12: 40–45.

Zhu, Z. and Woodcock, C.E. 2014. Continuous change detection and classification of land cover using all available Landsat data. Remote Sensing of Environment, 144: 152–171.

Last updated on Apr 08, 2026.

On this page
Edit on GitHub

This Page