Artificial intelligence and anticipatory action: a conversation - part 1

In a recent workshop organized by the Risk Knowledge Action Network, the ELLIS Unit Jena, the German Red Cross and the Malagasy Red Cross Society, members of the anticipatory action community met with artificial intelligence (AI) and climate researchers to discuss the use of AI in climate risk mitigation and anticipatory action. Two of the workshop’s facilitators, Karen Dall and Vitus Benson, continued the conversation afterwards. This blog captures the first part of their discussion, which focused on what AI is and how it might be relevant for these sectors.

Karen: Hi blog readers and hi Vitus. My name is Karen and I work for the German Red Cross as a technical manager for anticipatory action. In anticipatory action, we use forecasting and risk analysis to initiate and deliver assistance to communities before extreme events occur. This allows people to better protect their families and livelihoods from the impacts of these events and, ideally, to mitigate them.

Part of my work involves supporting National Red Cross and Red Crescent Societies, such as the Malagasy Red Cross Society in Madagascar, to develop pre-agreed plans for anticipatory action. These outline who does what and when, once a critical forecast threshold is reached. For example, when a flood is forecast in a specific region, the pre-agreed plans – called early action protocols – outline what the Malagasy Red Cross will do to protect the lives and livelihoods of the most vulnerable people.

Vitus: Hi all, I am Vitus, a PhD student at the Max Planck Institute for Biogeochemistry in Jena, Germany, and at ETH Zürich, Switzerland. I mainly work with AI, more specifically with large neural networks that I train on data about the Earth system, such as satellite images that can show the impacts of natural disasters.

We at the Max Planck Institute figured this research might also be helpful for anticipatory action. You also work a lot with data, right? But you are not really an AI expert, if I dare to say that! So, what do you understand about the term AI, and where do you think it could be useful for your work?

The AI hype

Karen: You are right, I am certainly not an AI expert and yet it is still part of my daily life, on top of the hype and media attention that it gets. When I think of AI, I think of ChatGPT, my email spam detector, Snapchat or Instagram filters, and generally models and algorithms that are learning by themselves.

This is probably also where AI is, and will be, most useful in my work. As mentioned, anticipatory action needs reliable forecast information to know when and where a hazard will strike in order to implement anticipatory actions. If AI can make forecasts more accurate and increase the lead time between a forecast and a hazard striking, this would be very valuable in my work.

During our workshop in Kigali, Rwanda, [held as part of the World Climate Research Programme’s Open Science Conference in October 2023] we not only focused on the potential of AI in forecasting, but also the broader question of how AI can enhance climate risk mitigation and anticipatory action. We brought together a group of AI researchers, data scientists and modellers with humanitarian practitioners to discuss the challenges and opportunities of AI for drought impact assessments and forecasting, and risk mapping and forecasting for floods and cyclones.

How does AI work?

Karen: Before diving into these topics, can you explain AI in simple words? Can you demystify the term for us non-experts? 

Vitus: Absolutely. AI, as you mentioned, is a big topic right now. I would argue the main success story of what we call AI today is supervised machine-learning – which people have actually been doing for over 30 years.

So, what is supervised machine-learning? Say you wanted an AI model that could translate text from one language to another, for example from German to English. What you would have done 10 years ago is collect labels; so, you would take a German sentence and then find the matching English translation. You would repeat this many times until you have a lot of samples. These constitute your input–output pairs, so AI gets a German sentence as an input and is supposed to predict the English translation as an output.

Of course, the AI model could not do so immediately; you had to train it. First, you let it output some random initial guess, and then you compared this prediction with the ground truth English translation. This first random prediction of the AI model was probably wrong, therefore you would tell it to change its output next time, so that the next output is closer to the ground truth. That’s what you did, many, many times and over many iterations and, ultimately, you would eventually have a model that can translate from German into English.

But this is not how ChatGPT and the current wave of AI models work. The idea nowadays is that collecting all those labels, all those input–output pairs, is a very tedious task, because you need a lot of data which results in a lot of work. Instead, what we do today is simply take all the data we already have; for instance, the whole internet, or all the images in Instagram. Or, to be domain-specific, all the satellite data that has been collected. And then we train an AI model by randomly dropping some of the data and predicting it. Basically, we train a model to fill gaps in the data.

Going back to our translation example, today we would give the AI model a German sentence and leave out the last word, and then say: AI model, predict the last word. If the model predicts the wrong word, we penalize it and update the algorithm so that the next time, it outputs a different word that is closer to the ground truth. This is what is now driving the success of AI, and the key aspect is this so-called self-supervised learning, which means that you can use much larger data sets.

This is one of the requirements of AI: you need data. And the more complex your problem is, the more data you will need if you want to solve it. This is fairly logical, but of course poses constraints. For instance, if you have a data set with a lot of biases in it, then those biases will most likely end up in your AI model. As AI researchers, we are working hard to mitigate all these problems.

AI for forecasting climate disasters

Karen: Can you give us an example where AI is currently used for humanitarian action, or more broadly in disaster risk reduction and mitigation?

Vitus: One particularly successful story, which was highlighted during the workshop, is forecasting. For instance, you can now do weather forecasts with AI. The way it works is that you train your AI model on all of the historical weather data. Now, the model knows the patterns in the weather, understands the circulation, the physics equations behind it, so you can use it to predict future weather, iteratively and one step ahead.

Currently, the best models are GraphCast from Google Deepmind, or PanguWeather from Huawei. These are at least as good as, and sometimes better than, the world-leading weather-forecasting systems, such as the European Centre for Medium-Range Weather Forecasts, and they only use a fraction of the supercomputing power. In particular, they are good at predicting the tracks of tropical cyclones, which is important for anticipatory action for cyclones.

I also want to plug one example of our work in the EarthNet initiative, in which we use satellite data and AI to forecast more impact-centric variables. On satellite images, you can often see the impact of a drought on the vegetation and ecosystems. So, the idea is to directly predict the future satellite images, so as to see the impacts of a drought that will happen in the future.