Nature is vital for our climate, economies, and lives. Forests are a key part of nature. They store carbon, control rain, reduce floods, and support most of the world's animal and plant life. Sadly, forests are being lost quickly. Last year, we lost huge areas of tropical forest. This is more than the year before. Losing forests is the biggest threat to animals and plants today.
For a long time, we used satellite data to measure forest loss. We worked with the World Resources Institute to map why forests are lost. This includes farming, logging, mining, and fires. These maps helped us create ways to protect forests. But these maps only show what happened in the past. Now, we need to look at the future.
We are happy to share “ForestCast: Forecasting Deforestation Risk at Scale with Deep Learning.” It also includes the first public dataset for training AI models to predict deforestation risk. This new approach helps us see future risks instead of just past losses. Old methods used maps of roads and people. These maps can become old quickly. Our new method uses only satellite data. It works the same everywhere and can be updated easily. It is as good as older methods. We are sharing all the data so others can use and improve our work.
Why Predicting Forest Loss Is Hard
Forest loss is caused by people. Many things like money, politics, and the environment play a part. It is driven by making products like beef, palm oil, and soy. It is also caused by fires, logging, new buildings, and mining. Predicting where and when forests will be lost is very hard.
The best current method uses information about these factors. It looks at maps of roads, money, and rules. This method has given good results in some areas. But it does not work well everywhere. The maps are often incomplete and need to be made for each area. Also, these maps get old fast. It is not clear when they will be updated.
A Scalable Satellite Method
To solve these problems, we use a method with only satellite data. We tested data from Landsat and Sentinel 2 satellites. We also used data showing past forest loss and the year it happened. We trained our model using satellite data that showed where forests were lost.
This satellite method is consistent. We can use the same method anywhere on Earth. This lets us compare different areas. Our model is also ready for the future. Satellite data will keep coming. We can use it to get new risk predictions and see how risks change over time.
To be accurate and work on a large scale, we made a special AI model. The model uses a whole area of satellite images. This helps it understand the whole landscape and recent forest loss. It then predicts risk for the entire area at once. This makes the model work for large regions.
Our model was as good as or better than methods using special data. It correctly predicted forest loss in different areas. It also predicted which parts of the areas were most likely to lose forests next.
Our AI model looks at satellite images over time and past forest loss to predict future risk.
We were surprised that the most important data was the simplest: the history of forest loss. A model using only this data was as accurate as models using all the satellite data. The history of forest loss shows how much forest was lost recently in different areas and how this is changing. It also shows where forest loss is happening.
To help everyone check our work, we are sharing the data we used. This lets AI experts check our results. They can learn more about why the model makes predictions. They can also build and compare better models for predicting forest loss.
Our work also shows how to use this method for forests all over the world. This includes forests in Africa and South America. It can also be used for forests in colder areas where loss is caused by different things, like fires.
Conclusion
Losing forests causes about 10% of global greenhouse gas emissions. It also harms most of the Earth's animal and plant life. Predicting where forests are at risk can help us stop emissions and protect nature.
Knowing the risk ahead of time helps governments, companies, and communities act early. They can stop forest loss before it happens. For example:
- A government can help groups in areas where forests might be lost.
- A company can check its supplies to make sure they do not cause forest loss.
- An indigenous group can protect the land most at risk.
So, a risk forecast is not about a future that will surely happen. It is a tool to help change the future. The goal is to share this information so people can help. They can send help to areas at risk. This will help keep these important forests standing. By using open data and smart AI, we are creating a strong new tool to protect nature.
Learn more about our AI and sustainability work by visiting Google Earth AI, Google Earth Engine, and AlphaEarth Foundations.