Bogor Ecosystems Workshop

Applying Google's Satellite Embeddings Classifier for rapidly mapping ecosystems in Indonesia. Bogor, Indonesia. July 1-4 2025.

Description

Ecosystem mapping has historically been a labour intensive and complex task that requires advanced geospatial skills, a deep knowledge of ecosystem ecology and an ability to activate a range of different data sets (field data, training data, validation data, existing maps) to support the mapping exercise.

In this exercise, we will learn to use the Satellite Embeddings Classifier (https://remap-app.org), a prototype tool developed by the Google Earth Engine team for providing quick and easy access to the new Satellite Embeddings dataset. The satellite embeddings are developed at 10-m spatial resolution and integrate a very broad range of satellite data into a single, analysis-ready dataset that has greatly reduced the resources required to develop ecosystem maps at any scale. 

Our task today is to apply a classification model to the embeddings to yield a map of Ecosystem Functional Groups for a particular region. 

Many thanks to Sean Askay and team at Google Earth Engine for developing the app that we are learning to use today.

Breakout session aims

Today, we will use the Satellite Embeddings Classifier to:

1. Understand the importance of carefully understanding the Global Ecosystem Typology, which forms the basis of the map classification scheme.

2. Learn the importance of carefully collected training data, which are confirmed occurrences of each ecosystem type in the classification scheme.

3. Explore how training data is collected and applied in a classification model, iterating our map to achieve the highest accuracy result.

4. Submit training data to the Global Ecosystems Atlas initiative and contribute to a global effort to provide improved data, tools and resources that support a better understanding ecosystems.

5. Provide the basis of a discussion about how training data could be collected for a national scale mapping effort in Indonesia.

Workshop materials

Global Ecosystems Atlas website (www.globalecosystemsatlas.org)

Global Ecosystems Typology website (www.global-ecosystems.org) 

Embedding Fields Classifier App (link)

Google Form for submitting your data (link)

Supporting materials

Introduction to Earth Observation and Remote Sensing (PDF)

A presentation about ecosystem mapping by A/Prof Nicholas Murray on Google Earth's youtube channel: https://www.youtube.com/watch?v=qzN0y884DOs&t=2338s 

Workshop instructions

Step 1. Open the Satellite Embeddings Classifier 

Navigate to the Satellite Embeddings Classifier application:
https://earthengine-ai.projects.earthengine.app/view/satellite-embeddings-classifier-geo 

Note: if you receive a firewall error, please simply wait for a few minutes and try again. This may take a few tries to resolve.

Step 2. Develop an understanding of ecosystems in your area of interest

Select an area of interest or a small geographic region for which you want to develop a map of ecosystems. We suggest a fairly small area (say 50 x 50km) to start with, and once you're familiar with the application you will be able to expand.

1. • Identify the Level 3 Ecosystem Functional Groups present in your area of interest. For example, a coastal area is likely to have some of the following ecosystem functional groups:

     1. • • MT1.3 Sandy Shorelines

          • MFT1.2 Intertidal forests and shrublands (mangroves)

          • MT1.2 Muddy Shorelines

          • MT1.1 Rocky Shorelines

          • T7.4 Urban and industrial ecosystems

          • T7.3 Plantations

          • M1.7 Subtidal sand beds

          • M1.1 Seagrass meadows

          • M1.3 Photic coral reefs

   • To develop a deep understanding of the Ecosystem Functional Groups you are mapping, we suggest you take the time to read the ecosystem descriptions of each of these ecosystem functional groups. Detailed descriptions of each Ecosystem Functional Group are available on the Global Ecosystem Typology website.

Step 3. Navigate to your area of interest in the Satellite Embeddings Classifier app

In the Satellite Embeddings Classifier App:

1. • Use the search bar to navigate to your area of interest.

   • Use the zoom button to situate your area of interest.

   • For the purpose of this workshop, we suggest a small study area to promote speed and accuracy during our short breakout session (suggest about 50km wide).

   • When you map window is centered on your area of interest you are ready to make a map. Note that when you develop your classification, the app will implement it across your full map window.

Step 4. Develop your map classes

Develop a classified map by training a model in the Satellite Embeddings Classifier App:

1. • To develop a classified map of the Ecosystem Functional Groups in your area of interest, you first need to develop a list of the ecosystem functional groups (which are your map classes) in the app. 

   • The + new class now enables you to make a label for a single ecosystem type. 

   • Name your new class (e.g. MT1.3 Sandy Shoreline) using the text box.

   • You can also choose a colour for your map using interactive colour selection box.

   • Now, we will aim to map all of the different ecosystem functional groups in your area. Do note you can increase the number of classes in your map as you choose and, in many cases, you may need to make an 'other' class or a 'deep water' class to enable you to map all of the ecosystems in your study area.

   • At this stage, you should have at least 2-10 classes in your map legend.

Step 5. Develop your training data

Use the + button to add training data to the app. The training data are simply point-records of each ecosystem type that you can confidently find in the map window. 

1. • Add 5-10 points per class. Use your knowledge of an area and its ecosystem types to find places where you are confident you know the ecosystem functional group and add a single point to that place. 

   • Once you have a few points per class, the classifier will begin to return a classified map. In this map, the application is estimating the most likely ecosystem type per pixel using the training data you've provided to it.

   • The application will return a map of ecosystems when you have provided training data for all of your classes.

   • To promote the development of an accurate map wsuggest at least 20 samples per class.

Step 6. Refine your map through iteration

Now that you have your first draft map, you can begin the process of refining your map. Typically, you would continue to work through improving your training data until you cannot make your map any more accurate.

1. • Continue to add training points, aiming for a well distributed set of training data that covers all of the ecosystem functional groups in your map. 

   • In the refining phase, it's often useful to focus your extra training data collection on places in your map that your first draft is not correct.

   • As you continue to add new training data, your map will slowly improve until you're unable to make any further improvements.

   • With this workflow, you've been able to develop a classified map of ecosystem functional groups for your area of interest!

Step 7.

When you feel like you've made a map of sufficient accuracy, please share your map back with the GEO/IUCN team. We'll be able to use the maps developed to track the outcomes of this workshop:

1. • To share your map, click the share button

   • Select "Share this map with the GEO/IUCN team"

   • Fill out the submission form --> Link to submission form

Notes and advice

Here are a few final tips to advance your map using the Satellite Embeddings Classifier:

1. • The map will only be as good as the training data, so try to be accurate with your training data.

   • You can try 'lumping' very similar classes. When classes such as two different types of wetland, or even oceanic versus lakes, it's often better to lump them into a single 'water' class. After making your map, we can then do a post-processing analysis to 'split' those classes back into Ecosystem Functional Groups.

   • Some ecosystems, such as tropical flooded forests, are extremely difficult to map using earth observation data alone. Therefore, in these situations it is often necessary to 'lump' similar forest types and then implement a post-processing analysis to split them back into particular ecosystem types.