Radiant MLHub Spotlight Q&A: Victor Faraggi

Accelerating climate change applications with machine learning models and remote sensing data.

Meet Victor Faraggi, our Community Voice for the second quarter of 2022. Victor is a Computer Science Engineering graduate from the University of Chile. There, he’s been part of the Teaching Staff of multiple courses, such as Algorithm Design and Analysis, Introduction to Data Mining, and Machine Learning. Currently, Victor is working towards his M.Sc. at the intersection of Graph Representation Learning and scientific discovery. The nature of his work sparks due to his interest in contributing his grain of salt to problems that can help tackle climate change.

“We cannot tackle all problems only with curated datasets. Climate-related events are complex and depend on multiple variables, so I hope that access to all free satellite data will improve. “ — Victor Faraggi

In this Q&A, we sat down with Victor to discuss climate change related models and the challenges to developing these, and what we could do to help scale sustainability projects.

“What Radiant MLHub has created sets an example of accessible and usable datasets. Easy access means that we can use them for fast prototyping ideas and even evaluate our data’s quality assurance processes.

You are currently enrolled in a Computer Science Engineering master’s degree program at the University of Chile with an interest in climate change and earth science models. What inspired you to pursue this field? Tell us about your machine learning journey.

What a great question! I think that, like many of the people working at the intersection of machine learning and climate change applications, my path was an unusual one.

The fact that both my parents have worked impactful societal jobs carved an early interest. In particular, as my mother has been working on adding ecological and climate views to Chilean public policies for at least a decade, I was aware of the importance of making a difference. However, during my first university years, I didn’t see how Computer Science could be able to contribute to that field, so I turned to Computer Security, where I could help with current privacy issues.

Then, back in 2019, during one of my first Machine Learning related courses, we were asked to present a paper from the latest ICLR conference. At that point, I stumbled onto the recently released CCAI paper and saw how ML could help by accelerating climate change applications. After that, I didn’t look back: I enrolled in every ML course that I could, searched and applied for every Chilean startup working at this intersection, participated in the ProjectX competition hosted by the University of Toronto, and later went for an academic exchange there.

There are many challenges with building ML applications using Earth observation data, such as (lack of) diversity and bias in data and the ability to scale research applications to real-world solutions. What challenges have you faced when you developed applications?

I think that it’s a multi-stage challenge.

First, the issues around the lack of data diversity are a known fact. For example, most of the sources of curated datasets, whether it’s satellite imagery or on-site measurements, are still sparse. This is a known limitation that can be tackled, on one side, by studying if transfer learning methods are sufficient for a given application. On the other side, data collection efforts are still necessary, and RadiantMLHub’s work is an example for the industry.

Second, I think that challenges unrelated to the data itself are even more challenging. Working with stakeholders is still one of the most complex parts of developing ML applications that tackle climate-related issues, especially in efforts where the immediate gains aren’t evident. It translates into a fraction of the work used by decision-makers.

As you know, Radiant Earth has various open-access training datasets available on Radiant MLHub. What does a data infrastructure like Radiant MLHub mean to you as a researcher and teacher in a developing country?

As I said before, what Radiant MLHub has created sets an example of accessible and usable datasets. Easy access means that we can use them for fast prototyping ideas and even evaluate our data’s quality assurance processes. Moreover, I think what’s best about Radiant MLHub isn’t only its data infrastructure: the tutorials, webinars, and workshops build towards a stronger community of practitioners applying AI for the social good.

You have worked with training data available on Radiant MLHub to build an application. Which specific training dataset(s) have you used, and for what purpose(s)?

Here I’d like to talk about two different projects so we can better highlight the advantages of using datasets from Radiant MLHub.

In the first project, we attempted to improve Deep Learning prediction of cyclone wind speeds. Prediction of tropical cyclone wind speeds is difficult due to the many variables and parameters involved. Current methods of estimating wind speeds in tropical cyclones are limited to inspection of satellite images of the storms and empirical comparison of various cyclone measurements dubbed the wind-pressure relationship. For this purpose, we trained our models with the Tropical Cyclone Wind Estimation Competition training dataset. At the time, the whole dataset wasn’t publicly released, so we contacted Radiant ML and their liaison from NASA. Communication went smoothly, and, with time, we received access to the rest of the dataset, documentation, and even code to manipulate the data efficiently. All in all, it was a great experience.

The second project went utterly different. This time we tackled landslides, from which countries like Chile will witness a rise in severity and occurrence due to climate change’s effect on precipitation, temperature, and drought. These events are a significant hazard to human life and society’s infrastructure, so we tried to work with the Chilean government to build an Early-Warning system using data from satellite images. To make our project usable by Chilean agencies, we were asked to use similar raw data that the agencies were using. This induced many problems in the data collection process, where we had to spend more than five times the time that we had expected to collect, clean, and organize the data. This fact hindered results and eventually led the project to a stop.

As a follow-up to the previous question: Can you share any insights on how the model(s) performed or how it was implemented?

Regarding the prediction of wind speeds, we noted that there had been recent efforts to use deep learning models to predict wind speeds from satellite images of tropical storms. We used a neural ODE network to capture the complex relationship among the relevant parameters. When incorporated with a convolutional neural network, our model can make more accurate wind speed predictions when compared to other traditional CNN-based models.

What is your hope for scalable machine learning on satellite data related to sustainability projects like climate change in Chile? What do you need to scale what you have built so far?

I really hope for two main improvements.

Firstly, we cannot tackle all problems only with curated datasets. Climate-related events are complex and depend on multiple variables, so I hope that access to all free satellite data will improve. These will allow stakeholders and AI practitioners from all over the world to work together efficiently.

Secondly, great awareness efforts are being made in North America, Europe, and even Africa, but Latin America, Asia, and Oceania are still lagging. Sadly, even Radiant MLHub is an example of this, where most of the datasets are from the three former continents. If we can improve this, non-academic sectors from the latter regions will begin to understand the capabilities of applying ML models on satellite data, and better solutions to local problems will lead.

From your perspective, what innovations in ML and spatial analysis can we expect in the next decade, and how do you think they might improve people’s lives?

Suppose we can resolve issues like the ones stated above. In that case, we will be able to bring solutions to a vast myriad of incoming problems due to lack of water access and increased climate variability. It will lead to AI-based improvements in natural resource management, early-warning systems, and general climate change readiness. Also, as the effects of climate change become more and more apparent, governments will improve their policies, and collaboration between the public, private and academic sectors will rise.

--

--

--

Helping the global development community navigate the Machine Learning and Earth observation marketplace and innovations taking place.

Recommended from Medium

Autoencoders in TensorFlow — part 3

Reconstruction: The Mighty Camera

A General Introduction to Machine Learning

A deep dive into the Google Inception Network

Application of Cantor’s Pairing Function in AI/ML Models

DepthWise Separable Convolution

Food Mnist Classification

Game AIs with Minimax and Monte Carlo Tree Search

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Radiant Earth Foundation

Radiant Earth Foundation

A non-profit on a mission to make insights from Earth observations and machine learning more accessible for global development organizations and communities.

More from Medium

Bridging the Land-Use Gap in Carbon Inventories with Satellite Imagery

Modifiable Areal Unit Problem — The Spatial Data Scientist’s nightmare

When satellite images look into the past: Earth observation & archaeology

How I built a wind map with MTS