Treetection

Inspiration

The inspiration behind "Treetection" stems from the urgent need to combat deforestation and its devastating effects on our planet. By harnessing the power of satellite imagery and advanced semantic segmentation techniques using UNet Architecture, we aim to provide a proactive solution for predicting and preventing deforestation.

What it does

"Treetection" utilizes UNet Architecture to segment satellite images and distinguish between forested and non-forested areas with high accuracy. This enables us to create predictive models for identifying regions at risk of deforestation, empowering environmental organizations and policymakers to take timely action to preserve our forests.

How we built it

"Treetection" was constructed using a UNet architecture, a powerful convolutional neural network specifically designed for semantic segmentation tasks. Our approach involved extensive preprocessing of satellite images to enhance their quality and extract relevant features. We then trained the UNet model on these preprocessed images, fine-tuning its parameters to achieve optimal performance in accurately segmenting forested and non-forested areas. Additionally, we developed a user-friendly interface to facilitate seamless interaction with the model and interpretation of results. We used annotated satellite images from the Amazon rain forest for training. Since time series imagery was unavailable, we simulated deforestation by creatively interpolating different images. Specifically, images are interpolated by gradually expanding the area of farmland into areas covered by forest. We show that such a feature could be useful for controlling the behaviour of suppliers.

Challenges we ran into

Throughout the development process, we encountered various challenges which was mainly related to data preprocessing. The Amazon dataset that we used contains satelite images as TIF files and it was hard to convert them to other images formats that can be inspected visually like png/jpg. We ended up reading the TIF files directly in the UNet and for visual inspection we utilized a Python code for that using the rasterio library.

Accomplishments that we're proud of

We're proud to have developed a robust and scalable solution that can predict deforestation hotspots using satellite imagery. Our main contribution lies in tying these predictions together in a system that allows an actor in the supply chain to monitor deforestation over time by their suppliers.

What we learned

Through the development of "Treetection," we gained valuable insights into satellite image analysis, deep learning techniques, and environmental conservation efforts.

What's next for Treetection

Moving forward, we plan to enhance the capabilities of "Treetection" by integrating additional environmental indicators and refining our predictive models.