Data Challenge Winners: Identifying African crop types using satellite imagery
A conversation with the First Place winners of the Radiant Earth Computer Vision for Crop Detection from Satellite Imagery data challenge.
In May, we announced the winners of the Radiant Earth Computer Vision for Crop Detection from Satellite Imagery data challenge, which took place in February and March 2020. A total of 440 data scientists signed up for the challenge, representing a wide range of educational backgrounds, institutions, and geographies. While five winners were selected, in this Q&A, we sat down with Karim Amer, the First Place Overall Winner of the Data challenge, and the First Place African Citizen winner, Femi Sotonwa. Our goal is to learn more about the people behind the top scores.
First Place Overall Winner, Karim Amer
“[Africa] has a lot of young and talented minds and knowledge is available to anyone passionate and ready to work hard. However, we need to identify our main priorities based on available resources.”
Meet Karim Amer, the First Place Overall Winner of the Crop Detection data challenge. Mr. Amer is a Computer Vision researcher at the Nile University in Egypt, applying deep learning to different problems in the last four years. He is also concurrently working on his Master’s thesis, which focuses on vision-based autonomous navigation systems for GPS denied environments.
Mr. Amer previously worked as a research assistant with the Ubiquitous and Visual Computing Group at Nile University, where he published several papers about satellite image analysis between 2016 and 2018. In 2019, he became a research intern at Siemens Healthineers Technology Center in Flanders, New Jersey, USA where he worked on the development of cutting-edge segmentation models that can be used in multiple clinical applications.
In addition to academic interests, Karim is fond of Japanese Anime and considers it a unique source of innovation.
In this Q&A, Mr. Amer talks to us about his journey to become a data scientist and ultimately winning the Radiant Earth crop detection data challenge.
Congratulations on winning the Radiant Earth Crop Detection Data Challenge! Tell us about yourself?
I am 27 years old with two siblings and live in Cairo, Egypt. At school, I excelled at Mathematics, and I wanted to understand how it could be used effectively in real life. That was why I decided to study Computer Engineering in college, where I learned about machine learning and how we can build intelligent machines that could improve our lives.
My undergraduate project was on the automation of the machine learning process, which can save data scientists and engineers a considerable amount of time when they do their work. As deep learning technology emerged, I decided to pursue my graduate studies in this innovative field. I wanted to apply what I am learning about real life problems and started participating in online machine learning competitions. I am currently a Kaggle Expert.
What did you think of the Data Challenge, and how did you approach it?
The competition was challenging and had room for creative ideas, which was interesting to me. The provided dataset was a time series of high-resolution multispectral satellite images of an agricultural area in Western Kenya. The images were acquired on 13 different days within five months. The objective was to classify the planted crop in some farms in this region.
After exploring the data, it became clear that a spatio-temporal neural network was the best solution. But I also realized that with no available pre-trained model on similar data, the hurdle would be to establish how to train the model on a small dataset (~3K training samples). My approach was to start with a simpler model and to gradually increase its complexity while adding more data augmentation to avoid overfitting.
My final model was a multi-layer Convolutional Neural Network (CNN) with a particular pooling layer. Seeing that some farms only occupied a couple of pixels and others held tens of pixels, I designed the model to take into consideration the area of input farms, followed by a multi-layer Gated Recurrent Network (GRU).
My key takeaway from this challenge is that neural networks are fantastic and have many degrees of freedom, which can be utilized for any structured data like images, text, time series, and graphs. It just needs some patience and consistency to get the best out of it as the tuning process usually takes some time before reaching the best configuration.
How did you learn about the Radiant Earth Data Challenge Competition?
I was excited to participate in the ICLR 2020 conference, which was going to be organized in Africa, my native continent. So, I searched for a data challenge at the conference and discovered Radiant Earth’s Computer Vision for Agriculture workshop and competition on the CV4GC website.
Were you familiar with the crop types used in the challenge?
As maize is one of the most common crops harvested in Egypt, I am familiar with its agriculture. Despite not knowing what cassava was, that it was a root vegetable similar to sweet potato, which is another widespread plant in Egypt. It was interesting to know that cassava is used for making wafers, while sweet potatoes are used for sweets.
How does it feel to beat out 440 top data scientists?
Firstly, I was ecstatic when I learned that I ranked first in the competition. My diligence, hard work, and intensive research were rewarded by being considered the best in that competition. Yet, I was also humbled and overwhelmed by having my research acknowledged by experts, as well as peers. I thoroughly appreciate the work submitted by some of the finest minds in Africa and worldwide. Finally, I feel that the prize I received is an incentive that will encourage me to conduct further research.
What do you think are the opportunities for ML research in Africa? And what role do you think ML techniques can play in addressing development challenges in Africa?
I believe we can do cutting-edge ML research in Africa because we have a lot of young and talented minds and knowledge is available to anyone passionate and ready to work hard. However, we need to identify our main priorities based on available resources. Working on innovative research problems like few-shot learning, self-supervised learning, and fast convergence optimization can have a positive impact that we can utilize in Africa.
Furthermore, it is essential to link research with our current problems to help African people have a better life starting with their basic needs like food, health, and education. For instance, we can make use of ML systems for plant disease detection and precise farming to increase food productivity.
First Place African Citizen Winner, Femi Sotonwa
“It was my first time working with remote sensing data, which made me appreciate the field and its capabilities.”
Meet Femi Sotonwa, a self-taught data scientist from Nigeria. An accountant by training, Mr. Sotonwa participates in data challenge competitions in his free time. Fascinated with computer science and technology, even while studying accounting, he taught himself to software development until he discovered the practical application of data science to real-world problems. Mr. Sotonwa has participated in five competitions thus far. He says, “data science provides a means to get hands-on experience on various kinds of social problems, as well as data, without the need for exclusive access.”
Where are you from, and what’s your educational background?
I hail from Nigeria. I am a graduate with a Bachelor’s degree in Accounting from Babcock University, Nigeria.
What did you think of the Data Challenge, and how did you approach it?
The challenge was an interesting one. It was my first time working with remote sensing data, which made me appreciate the field and its capabilities. I approached the challenge by looking at the problem from two perspectives, each with a similar modeling process. Implementing a solution from two approaches, and later combining them gave a boost to the performance. Dealing with the inherent problem of class imbalance by using class weight also helped.
While the competition was open globally, all three winners came from the African continent, which indicates how tech hubs play an essential part in innovating businesses on the continent with no signs of it slowing down. What are some of the opportunities and challenges facing data scientists in Africa?
Various opportunities exist that need further exploration. Some of the interesting ones include Natural Language Processing and Computer Vision. State-of-the-Art implementations in these areas usually involve Deep Learning (DL) using Neural Network architectures. DL is relatively costly to implement, depending on the scale of the problem at hand. Challenges — at least from where I come from — range from reliable standby electrical power, relatively affordable internet data, and the availability of resource persons, amongst others. Tech Hubs are beneficial in curbing some of these challenges, but it remains a barrier to many.