Radiant MLHub Spotlight Q&A: Gedeon Muhawenayo

Building machine learning models with open training data for precision agriculture and flood detection in Rwanda.

Radiant Earth Foundation
Oct 5 · 6 min read

Our Community Voice for this quarter is Gedeon Muhawenayo, a machine learning research engineer at the Rwanda Space Agency working on machine learning for satellite and aerial image processing. Rwanda launched its first CubeSat satellite, RWASAT-1, on November 20, 2019, to monitor agriculture. Gedeon is currently on secondment at Inria, the French National Research Institute for Digital Science and Technology working on machine learning, optimization, sparse coding, and hyperspectral data processing projects.

Gedeon holds a masters degree in Machine Intelligence from the African Institute for Mathematical Sciences (AIMS) in Ghana. He researched deep learning approaches for compressed object detection without a loss in performance. Passionate about applying machine learning to large datasets to create viable solutions, he discovered Radiant MLHub through Radiant Earth’s first virtual training bootcamp focused on using machine learning to satellite data. The bootcamp ran from May 3–14, 2021, and was sponsored by a grant from the GIZ FAIR Forward- Artificial Intelligence for all initiative, which is implemented by the Deutsche Gesellschaft für Internationale Zusammenarbeit (GIZ) on behalf of the German Federal Ministry for Economic Cooperation and Development (BMZ).

Today, Gedeon is an avid user of the open machine learning training datasets available on Radiant MLHub, both personally and as part of his work at the Rwanda Space Agency. In this Q&A, Gedeon talks to us about how he — and other colleagues — are building models for precision agriculture and flood detection using Radiant MLHub training data.

Tell us about your machine learning journey. Your bachelors’ degree is in Electronics and Telecommunication Engineering from the University of Rwanda. How did you become interested in the intersection of machine learning and satellite imagery? What inspired you to pursue this field?

Doing my bachelor’s in applied sciences, particularly electronics and telecommunication engineering, has exposed me to different sub-fields, including satellite communications and applications of small satellites. I was fortunate to be among the few who got trained in nano-satellites and satellite mission design applications. This training, in addition to the exposure to satellite-based applications, has boosted my eagerness for the usage of satellite data, specifically satellite images. The ability to capture information from the sky and assess changes to the environment was mesmerizing for me. Since then, I have decided to grow my career in Earth observation data processing.

You work with various training datasets available on Radiant MLHub to build models for precision agriculture and flood detection in Rwanda. First, can you talk about why these thematic areas are essential in Rwanda and which specific datasets you are using?

Rwanda is prone to a wide range of natural hazards, including landslides, floods, earthquakes, and windstorms, which strain the most vulnerable communities and take a considerable toll in terms of economic losses and human lives. These natural disasters are likely to get more frequent and intense if you add on climate change factors.

Furthermore, agriculture is a major economic sector for the people of Rwanda, so I think working on a project that would optimize inputs (fertilizers, irrigation, types of crops, etc.) is necessary for food security. Applying machine learning on EO data can support the ways to increase harvests for the Rwandan community.

Various datasets hosted on the Radiant MLHub make it easy to build predictive systems for disaster and agriculture applications. One reason is that the datasets that Radiant provides are machine learning ready. The training datasets are high-quality, structured and one does not need to clean and normalize the data. Radiant’s training datasets are helpful, especially when you don’t have enough labeled training data for your area of interest. You can use it to train your model to make decisions based upon the data and then to transfer the knowledge to your specific objective.

I used the crop type classification datasets from different countries that are available on Radiant MLHub for the agriculture project I am working on, then transferred the knowledge to our specific domain.

We have used the SEN12-FLOOD SAR and multispectral dataset for flood detection training data to build a flood application. Despite the difficulties of SAR data processing and limited ground truths, Radiant’s data are well prepared.

As a follow-up to the previous question: Can you share some details on the applications you are working on and any insights on the preliminary research or how the model(s) is performing?

Let me start by highlighting the problems that the applications would address.

It’s hard to estimate crop yield and to take administrative decisions or implement new policies accordingly in many African countries. It was challenging to address this problem using predictive approaches and data analytics tools mainly due to the lack of high-quality training data. Now with public EO training datasets like those Radiant MLHub hosts, one can change this trajectory.

We are currently building an application that various stakeholders could use to estimate crop yield, crop diseases, and fertilizer maps, and conducting a crop life cycle analysis for Africa. To inform our design, we read many recent state-of-the-art research papers in remote sensing, geospatial analysis, and machine learning.

The whole system we’re building consists of subsystems, where AI powers the core module. We have included many containerized models to run the application. For instance, the baseline crop type classification model performs pretty well (over 70%), but this is not the desired performance, so we need to improve it with a data-centric approach. By focusing on the data tuning, we could improve the model performance.

Apart from the open access, what is the appeal of using Radiant MLHub data for you and your colleagues? What are some specific benefits that you are receiving using Radiant MLHub?

Besides easy access to the training datasets, Radiant MLHub has good documentation and start codes. These are very useful to me and anyone else because one does not have to begin from scratch. Furthermore, the Radiant Earth technology team is approachable and easy to reach, either directly on Slack or via email. They tend to respond pretty quickly.

I was also fortunate to participate in the ML4EO training organized by Radiant Earth. This knowledge transfer has allowed me to learn more from their experts.

You’ve mentioned in previous communications that you are also using Radiant MLHub data to expand your knowledge and scope using the data for commercial purposes. Can you share your vision and what value it will add to society in Rwanda?

I do use Radiant MLHub datasets for my professional development, but also to train other Rwandans. I can do this because of the starter notebooks and tutorials available on the Radiant Github repository. Establishing my startup is one of my main career goals, together with other colleagues who are on board with my vision. We are ideating a mobile web app that would be an intelligence system powered by well-trained ML models, and the primary datasets are remote sensing images. So far, Radiant MLHub datasets are our best option. We believe in the idea of object recognition from high height.

You’ve completed the ML4EO Training of Trainers Bootcamp and are currently busy with the last requirement of training five people. What insights have you learned as a trainer versus being a student, and how does teaching others benefit you?

I have learned that being a trainer is the best way to learn and master things! It exposes you to unheard and unexpected questions, which require reading and sometimes asking more experienced people. Personally, teaching others has boosted my skills in remote sensing, geospatial data analytics, communication, and applying machine learning on EO.

What is your hope for scalable machine learning on satellite data, related to food security and natural disasters like floods in Rwanda? How can what you build be used to scale in the region or even in the rest of Africa?

African countries have common agricultural systems, largely small-scale and subsistence farming, and have similar climate conditions and landscapes (albeit at a regional level). Therefore, it is entirely possible that what works in one African country could also work in other countries on the continent. So I believe a system working well in Rwanda could also work in other African countries with slight modifications and a kind of domain adaptation.

From your perspective, what innovations in machine learning and satellite data analysis can we expect in the next decade, and how do you think it will improve lives in Africa?

With more access to the data, intelligent model repositories, and more Africans in the field of AI to build local solutions, I think machine learning and satellite data will improve food security, disaster management, and urban planning very soon.

This would save more lives and address some of the Sustainable Development Goals (No poverty, Zero hunger, Life on Land, etc. ), leading to the sustainable development of the whole continent.

Radiant Earth Insights

Earth Imagery for Impact