Melbourne Datathon 2019 1st place solution knowledge share

Background: Our team Klora win the Melbourne Datathon 2019 data2App 1st place solution. I would like to share some useful tips and experience that maybe useful to help you to achieve a better score in your next datathon or data2app competition. Before you read this blog I suggest you to go through the background knowledge under this link: https://medium.com/satellite-intelligence

1. Tech Stack

This part I would like to share the tools and frameworks we have been used for this datathon.

Dash(plotly):

Dash: React + Plotly (frontend), Flask (backend)

Here are the reasons why we choose dash as our development framework.

• A Dashboard is a good way to present the map and indexes.
• Most of the team members know python(more people to contribute)
• less code == save time

There are also some drawbacks to using the dash framework.

• The official document is not good
• The performance is not good

Before you choose Dash I suggest you have a look at this page https://dash-gallery.plotly.host/Portal/ These are the example applications.

Cloud and Tools:

Cloud Architect

The above image is the GCP services and tools we have been used.

Google Earth Engine:

To deal with the satellite imagery

2. Project Analysis(less is more)

Klora

Business analysis is important for this data2app competition. In the first four weeks, rather than spending too much time on coding we did spend a lot of time analyzing who is our potential customers and what value we can provide to them.

Focus on core customers:

The version one of Klora is very complex as we would like to provide services for government, banks, futures, farmers…

Then the problem is we cannot present so much information within 10 minutes presentation and 5 minutes video so that we delete some pages and focus on our core customer banks and farmers.

Focus on usability:

The version one of Klora used to have 40, 000,000 data points on the map. We allow customers to select sugarcane areas by themselves but then we realize that this is not user-friendly. So that our geospatial expert creates some pre-defined areas for banks and farmers to choose.

Apart from the above two tips, I would like to say that we find at different stages you need to use different prediction models. For example, at an early stage, you cannot rely on satellite imagery as you can see in the image below. You cannot extra too much information related to sugarcane yield.

We divide the sugarcane growing process into three stages(early stage, middle stage, later stage) at different stages we have different weight in terms of the feature selection.

For more information about the machine learning part, I suggest you read this article written by my teammate: https://medium.com/@wangminzheng1121/klora-1st-place-solution-for-2019-melbourne-datathon-9623a497a401

3. Project Management

If you have a big team(more than 5 people), team management is very important. Here are the management tools we have been used for this competition.

Nuclino:

This is a free platform for documentation and knowledge share. You can see these are the documents we have been created for this datathon. It turns out to be useful. If your team is not on the same page then you cannot expect them to be productive.

Trello:

We also follow the agile principles and use trello to manage all the tickets.

4. Conclusion

I used to join some other data competitions but I didn’t achieve a good score since most of the datathon need some domain knowledge. Do not treat Datathon as a pure machine learning task. If your team does not have enough domain knowledge then it is less likely to win.

Here are the two tips I would like to share. I hope it will help you perform well in your next datathon or data2app competition.

• Go to local data science meetups and communities to find teammates.
• At least one team member who has enough domain knowledge.

In the end, I would like to give a big thank to all my teammates. Without such a good team we cannot make it happen.