ForestGuard

Inspiration

The rampant issue of deforestation globally not only alters landscapes but also impacts biodiversity, climate change, and indigenous communities. We were inspired by the urgent need for actionable intelligence on this front.

What it does

Our project, "ForestGuard" uses satellite imagery to identify and classify areas of deforestation. It employs a two-tiered machine learning model: the first layer conducts binary segmentation, distinguishing forested areas from non-forested ones. Subsequently, a more detailed segmentation classifies the identified non-forested areas into various semantic land cover types such as agricultural land, urban areas, and water bodies. This detailed classification helps in understanding the nature and impact of deforestation in different regions.

How we built it

The platform was built using Fastai for image processing and model development. We utilized satellite images from DeepGlobe and then applied convolutional neural networks (CNNs) for the segmentation tasks. The models were trained and validated using a dataset that included annotated satellite images indicating various types of land cover.

Challenges we ran into

One of the major challenges was the accurate classification of mixed pixels, which are common in satellite imagery due to varying resolutions. Differentiating between natural forest loss and human-induced deforestation also posed a significant challenge due to the subtle differences in the satellite images. Additionally, dealing with a vast amount of data required efficient data handling and processing capabilities to ensure timely analysis.

Accomplishments that we're proud of

We successfully developed a prototype that can distinguish between forested and non-forested areas with high accuracy. Moreover, our semantic segmentation model can classify the type of land cover with significant precision, which is crucial for understanding the context of deforestation.

What we learned

Throughout this project, we gained insights into the complexities of ecological monitoring using satellite imagery. We learned about the importance of data quality, the challenges of working with mixed-resolution data, and the potential of machine learning in environmental science. This project also improved our skills in image processing, model tuning, and handling large datasets effectively.

What's next for Deforestation classification

Moving forward, we plan to enhance our models' accuracy by incorporating more granular data and adding more auxiliary data. Adopting a time-series approach may also greatly increase the accuracy of driver classification for deforestation events. We aim to collaborate with environmental organizations and government bodies to make it a practical tool for combating deforestation globally. The ultimate goal is to develop a fully automated system that can not only detect but also predict potential deforestation activities before they occur.