Apply responsible AI principles when building remote sensing datasets, such as Dynamic World

1. Before you begin

Dynamic World is a global land-use land-cover (LULC) dataset that's updated nearly in real time. It provides a 10-meter resolution because it uses the Sentinel-2 constellation by the European Space Agency (ESA) and the Copernicus program, which generates an updated global picture of the Earth's surface every 2-to-5 days. It's produced with a deep learning model, Vertex AI, and Earth Engine, which is a geospatial-analysis platform. It helps scientists, researchers, and developers detect changes, map trends, and quantify differences on the Earth's surface.

Earth Engine is a tool that seeks to advance scientific knowledge of planetary processes to further environmental and social benefit. Its commercial use, in preview with Google Cloud, states that its primary purpose is to understand and tackle critical sustainability and climate issues, such as deforestation, water management, and sustainable land use.

Note: The clarity of language about Dynamic World and Earth Engine is a common responsible AI approach to safeguard against potential use for any reason other than the intended purpose of societal benefit.

In this codelab, you apply responsible practices in the generation of machine learning-derived datasets, such as Dynamic World.

What you'll learn

• How to apply responsible AI practices in remote-sensing apps.
• How to visually interpret Dynamic World predictions over space and time.
• How to reflect on ethical implications and downstream societal implications.

What you'll need

• Familiarity with how to navigate the Earth Engine Apps.
• Google Chrome, which works best with the Earth Engine Apps.

2. Open the Dynamic World Earth Engine app

The Dynamic World dataset contains per-pixel predictions of nine LULC classes as described in this table. Land use describes how humans use the land, whereas land cover describes the physical material on the surface of the Earth.

LULC type	Description
Water	Permanent and seasonal water bodies
Trees	Primary and secondary forests, and large-scale plantations
Grass	Natural grasslands, livestock pastures, and parks
Flooded vegetation	Mangroves and other inundated ecosystems
Crops	Row crops and paddy crops
Shrub and scrub	Sparse-to-dense open vegetation that consists of shrubs
Built area	Low- and high-density buildings, roads, and urban open space
Bare ground	Deserts and exposed rock
Snow and ice	Permanent and seasonal snow cover

Images in this dataset include 10 bands: nine bands with estimated probabilities for each of the nine LULC classes and a class label band that indicates the class with the largest estimated probability. These distinct properties let users do multitemporal analysis and create custom products suited to their needs.

Note: This use of predictions in Dynamic World went through an ethics review in which it was evaluated for safety, scientific rigor, fairness, potential misuse, social benefit, and, in this case, alignment with Google's AI Principles. If applicable, you can set up checkpoints in your ML project to evaluate how the technical decisions you make align with your organization's mission statement, values, or ethical charter.

To open the Dynamic World Earth Engine App:

1. Open another tab in your browser,
2. Navigate to Dynamic World.

The map displays Top1 mode mosaic with hillshade at high zoom levels.

3. Click the map to view a chart that displays class probabilities over time.
4. Click a datapoint on the chart to load a specific prediction and its Sentinel-2 image.
5. In the Layers pane, adjust the layers and visibility.

3. Explore Dynamic World predictions in Brazil

Dynamic World includes predictions across all nine LULC classes and displays a single Top1 label for each pixel. The real world is a mix of many LULC classes at any spatial resolution. In this section, you visually interpret predictions over space and time.

To explore Dynamic World predictions in Brazil:

1. In your browser, navigate to the Dynamic World EE app.
2. In the Search places field, enter Sete de Setembro, Brazil.

You should see Dynamic World compute data for the indigenous territory of the Paiter Surui people, who live in the Sete de Setembro region of Rondônia, Brazil.

3. In the Dynamic World pane, enter 2016-01-01 in the Start date field and 2022-01-01 in the End date field, and then click Update.

The map may take a few minutes to load because it computes in real time.

4. Click anywhere on the map to generate a chart in real time.

You can zoom further in for a faster response.

5. Examine the chart in which class probabilities are on the y axis and time is on the x axis.

In this example, the Top1 class prediction from Jan 1, 2016 to July 27, 2017 is trees, which varies from a probability of 0.684 to 0.755. By the time the next Sentinel-2 scene is captured on Sept. 5, 2017, the prediction is no longer trees and the signal becomes much noisier.

6. Click to enlarge the chart.

Note: When you want to responsibly generate an ML-derived dataset, scenario-plan for how the dataset could be adapted for unintended and even harmful uses. Ask questions such as, "Are certain individuals, markets, regions, user segments, or populations at higher risk of negative consequences from the use of this dataset versus others?".

4. Congratulations

You learned how to apply responsible AI principles in the generation of ML-derived datasets, such as Dynamic World.

Learn more

• Responsible AI practices
• Introduction to Dynamic World (Part 1): Visualization and Creating Composites