Meet Stella Kimani (Plato), Azer Ksouri (ASSAZZIN), and Mohammad Alasawedah (masawdah), winners of the Microsoft Rice Disease Classification Challenge. The challenge attracted more than 800 participants, all vying for a $3000 prize pool. The objective of this challenge was to build a model to identify disease in pictures of Egyptian rice. Read on to hear how our winners beat the competition!

1st Prize: Stella Kimani, from Kenya

Please introduce yourself.

I am Stella Njeri Kimani (Plato) from Nairobi, Kenya and I work as a consultant at PWC.

Tell us a bit about your solution, and the approach you took.

I used both the RGB and RGN images for training - a balanced mix of two RGN images and two RGB images, together with Mixup greatly reduced overfitting of the model. The model was trained on a 10-fold stratified cross validation strategy to ensure class imbalance was maintained across the folds. The model was trained on the first five folds.

Other techniques:

• I used early stopping during experimentation to get the cut-off epochs, where there was no further performance improvement past certain epochs.
• I used extensive image augmentation to further curb overfitting.
• To accommodate the huge model, I used NVIDIA’s Automatic Mixed Precision to fit the model and data in memory.
• Pseudo-labelling with predictions over or equal to a probability of 0.999.
• WandB logging tool was used to track experimentation results.

How do you prepare for a challenge?

Actively doing research, and I read up on new models used on similar challenges.

Words of encouragement for others, or advice that has helped you?

It really helps to understand the challenge and the metric of evaluation before starting to build a model; also excellent data exploration and visualisation makes a major difference. Explore different models, for example ensemble methods where you use multiple models for prediction.

What do you like about Zindi?

I love the Zindi community! The platform also recognises the presence of beginners and supports them by providing beginner-friendly challenges including starter notebooks, which boosts their confidence.

2nd Prize: Azer Ksouri, from Tunisia

Please introduce yourself.

My name is Azer Ksouri, I am a newly-graduated engineer from Tunisia's SUPCOM. I have two years of experience in machine learning and data science.

Tell us a bit about your solution, and the approach you took.

I joined this challenge two weeks before the end, and was shocked by my baseline notebook's poor performance. As there was a strong correlation between Local CV Scores and Public LB Scores, and as I observed newcomers to computer vision climbing the LB with high scores, I was convinced that I needed to alter my strategy for this competition.

After a week of exploration, I understood that:

1. Images are small (500x400), so using large image sizes for cropping (as a data augmentation) would not be useful in this challenge.
2. Based on the last hint, I used 2 resizing configs: one to resize images and one for cropping.

This was a game-changer. My approach after this was to use heavy augmentations:

• color augmentations such like equalize, CLAHE, hue/saturation, and color jitter
• blur augmentations
• brightness augmentations
• flip augmentations
• rotations/translations

What set your winning solution apart from others?

Transformers models out-performed all others. Swin-Transformer (large) was really amazing in this competition!

How do you prepare for a challenge?

I have been preparing by participating in many Zindi challenges - this helped me understand how to approach different challenges and come up with the right solutions to them.

Words of encouragement for others, or advice that has helped you?

In any competition, start by going through the starter notebook (if it's available), understand the problem, and search for solutions on similar problems. This will help you ideate solutions in reference to the current problem outlined in the competition. The learning process takes time - be consistent and make use of online resources.

What do you like about Zindi?

Zindi is a good place to start accelerating your career as a data practitioner. Zindi also supports inter-community learning and growth, where people from different communities are able to post solutions that we can learn from. I have also sharpened my coding skills through Zindi. The competition data is usually not huge (in terms of volume), hence motivating new competitors and learners to indulge in various hackathons.

I like the new feature that shows public leaderboard and private leaderboard at the same time. You should improve this feature by adding the change of rank from the public leaderboard. This will be a good indicator for competitors’ robustness! :D

3rd Prize: Team DiseaseFighters, Mohammad Alasawedah, from Jordan

Please introduce yourself.

My name is Mohammad Alasawedah (masawdah) from Jordan. I'm working as an Earth Observation researcher at Eurac Research in Italy.

Tell us a bit about your solution, and the approach you took.

We started testing different models on both RGB and RGN separately, then we realized that the RGN images are not as helpful, as they need more time to converge because of their large size. Therefore, we decided to use just RGB images. We used a Swin-Transformer with different augmentation like resize, flipping and lightning.

What set your winning solution apart from others?

Training the model on different datasets using stratified k-folds, then blending the results made a big improvement.

How do you prepare for a challenge?

First we read about rice diseases to understand the problem. Then, we looked for related work, and read some papers about classification of rice diseases.

Words of encouragement for others, or advice that has helped you?

Don’t complicate things and try to do some fancy stuff! Always start simple, then try to improve.

What do you like about Zindi?

It’s a great platform to learn and make a difference in the world, by contributing to real-world solutions on the African continent and beyond.

To apply some of the lessons from our winners, take a look at some similar challenges currently on Zindi.