Satellite Applications Catapult
IDENTIFYING ILLEGAL MINING IN COLOMBIA
FUNDED BY
REQUIREMENT
Can we detect illegal gold-mining activities and predict for the Colombian police where they might be happening?
CHALLENGE
Noisy data and video feeds from one of the most cloud covered areas on earth were the major issues in the project that tries to detect change of activities over time.
SOLUTION
Out solution was an end-to-end machine learning model, coupled with a user interface targeted to gold mining experts that could evaluate the results and use them for detecting and predicting illegal activities.
Illegal gold mining in Columbia
Illegal Gold mining is thought to make up 80% of gold mined in Colombia. It damages the environment and local communities, and is a key source of fnance for criminal organisations.
The Columbian Police have sought ways to stop this crime. The UK Space Agency believed its data could help.
Separating the chaf from the grain in noisy data
There was a wide range of data – satellite data, geological maps, police reports – but each data set was quite limited, and a lot of it was noisy and messy. Turning it into useful insights was a challenge. That’s what Decision Lab were called upon to do in this Satellite Applications Catapult led project.
Our second challenge was on the human side. How could we get the police to trust the system and use it to increase their arrest and investigation efficiency?
A collaboration between engineers and mathematicians
We needed an approach that combined the diverse and sometimes conficting data to provide a common picture of events. But we knew that the sparseness of the data and its noisiness would mean that any predictions could have a high error rate – so we hit on the idea of incorporating the expert in the solution, so it could learn as it was used and the user could start to build trust in the system.
Developing a common methodology for the end user
We used data analysis technics to process the data and identify the connections between them. We used Machine Learning to build a robust model to predict in real time whether new information is telling us there is a crime. We present these assessments to the user and invite them to provide feedback – does the expert think they are right or wrong. The AI model then learn from this expert human judgement and gets better over time. It meant the expert user feels they are part of the process and they see how the system learns from them and build up their trust.
An exciting journey ahead for this new transformation
The tool that we managed to build during this project acts as an active learning loop between the humans and the model – the expert evaluates the model results and the system learns over time. This self-improving Feedback AI can be used to solve many business problems in many different industries where a human expert brings insight that is not evident from the data.
Let’s collaborate
Send us an email, to discuss a new project.
We’re a team of innovators who are excited about unique ideas and help companies to create amazing solutions.