Real world AI: using Custom Vision for image recognition to improve data quality for environmental studies

Real world AI: using Custom Vision for image recognition to improve data quality for environmental studies

On september 5th I got to talk on Global AI Night hosted by Microsoft NL about using the Microsoft Azure Custom Vision service in a real world AI case: species recognition in a large Dutch angler community platform. Goal: to improve the quality of the data that is used both internally on the platform and externally for environmental research studies.

Below a short introduction on the topic with links to a full video registration of the talk and my sheets.

I will post a more extended version of this article with a follow-up on training the model soon, so if you're interested, let's connect on LinkedIn.

Introduction

This real world AI case is all about the app "Vangsten" that is used by anglers in The Netherlands to report and share their catches. Their collective data (30k members, 600k+ catch records) can be used in environmental research studies. Combined with other data, researchers can analyse changes in the distribution of species in the country looking for long-term effects of important events in our surroundings, like changing water quality, water pollution, changes in (maintenance) policies and the creation or removal of migration barriers in water systems.

Custom Vision Goals: UX and Data Quality

The value of this platform boils down to one thing: community value, and that value depends highly on data quality. This is a social network and members want the data to be of great quality and free of errors. Also, to be used in research studies the data has to be correct and validated.

Being a citizen science instrument, a lot depends on the user filling in the report form. Since nowadays a picture gets added to almost every catch record, image recognition could analyse this picture during entry and dynamically help the user select the right species.

Furthermore, a strong and smart enough model would not only help improve the quality of new (future) entries, but can also be helpful in retrospect, by projecting it on existing (older) catch records to get an impression of the likelyness that these records have the right species selected.

The full video registration of the talk is below, including my approach in training and optimizing the Custom Vision model, results of the first 1.000 predictions and learnings for improvement. You can also download my sheets from the talk.

Let's get connected for follow up!

If you're interested on the follow-up -training and improving the model- or have any questions/comments let's connect on LinkedIn.