Blog 2 | Cognitive Deforestation Prevention | Deloitte Impact Foundation


Blog 2: Predicting deforestation

Blog 2/4 from the series about the Deloitte Impact Foundation initiative ‘Cognitive Deforestation Prevention’ that aims to prevent illegal deforestation by building an artificial intelligence solution that predicts where illegal deforestation will happen.

Joanne Lijbers works within the Artificial Intelligence department at Deloitte. As a data scientist, Joanne is in charge of the modelling part of the solution. In her blog she explains how the Early Warning System predicts the areas of potential deforestation, how these areas are prioritized and how the model’s accuracy is ensured.

Gaining impactful insights

Joanne: “For the past few years I have participated in several Deloitte Impact Foundation projects. I really enjoyed sharing my knowledge and learning from each other. My involvement in these projects however was always short-lived, as I mainly joined smaller projects that lasted a few hours or a day. That I now got the chance to work on a WWF-project which is supported by the Deloitte Impact Foundation on a full-time basis and in my specific area of expertise, is something I am really grateful for! 

Being a consultant at the Analytics & Cognitive department for almost 4 years now, I have helped several large companies with gaining strategic insights from their data. I think it is extremely cool that we are now helping the WWF with the exact same thing: using (satellite) data to gain insights in the subject of deforestation.

Prioritizing important predictions

As Mark explained in his blog the goal of the project is to predict illegal deforestation before it happens. So far we focused on predictions 6 months ahead. Using labelled satellite images combined with other data such as the location of forest fires and distance to population, we predict the location and likelihood of deforestation on a 480 by 480 meter level. The predictions are displayed in an user interface. A prioritization framework instantly lets the users know which predictions need action first. For example: The solution which is currently active in Kalimantan prioritizes predictions on whether or not specific species live in the area (e.g. orangutang; clouded leopard) and how much carbon is stored in the trees. 

Cognitive Model Early Warning System

Accurate predictions

In my role as data scientist I am in charge of the modelling part of the solution. For the past few months I have worked on improving the predictive algorithm, to result in accurate predictions. The improvements have been focused on what we call ‘user’s accuracy’ and ‘detection rate’, also known as precision and recall. The user’s accuracy shows the fraction of predictions that turn out to be true 6 months later, and the detection rate summarizes the amount of actual deforestation incidents we were able to predict. By testing different model parameters, input data sets and training methods, we were able to improve the detection rate by over 30 percent while remaining a high user’s accuracy.

User’s accuracy is user’s trust

Improving these performance measures is not only important for the sake of accuracy, but is also important to building trust with the end user. If the end user knows most of the predictions will turn out to be true, he/she is more likely to keep using the solution and more forest will be saved. With this reason in mind we have worked together closely with the end-user, ensuring they understand the solution and provide their input on both the modelling part as well as on the design of the user interface.

I really like that this project shows that as Deloitte, we can make an impact on society by using our own expertise, such as state-of-the-art data science and cloud computing. Personally, I’m proud to see that collaboration on these kinds of projects is at the core of our strategy!”

Did you find this useful?