This project represents an innovative application of deep learning to the critical environmental issue of forest damage. Utilizing aerophotos of forests and the U-net architecture, the study analyzes masked pairs of images with forest damages, providing insights into affected areas.
Key Components:
- Model Training: The U-net model was meticulously trained for 47 epochs at 49 seconds per epoch on a GPU P100.
- Data Source: The dataset was kindly provided by HackatonExpert Group and IPT faculty of Kyiv Polytechnic University, containing images only from November, leading to a biased dataset.
- Rapid Development: The project was developed in a Solo effort in just one day, demonstrating an agile approach to problem-solving.
- Recognition: The effort culminated in 1st place by jury points, reflecting the quality and innovation of the work.
Project Workflow:
- Image Masking: Utilizing the U-net model to create masks from pairs of images depicting forest damage.
- Damage Analysis: Applying the trained model to analyze remaining images, calculating percentages and square kilometers of damaged forests, both in general and by specific forest type.
- Visualization and Analysis: The next stage involves visualizing the results and conducting an in-depth analysis, providing valuable environmental insights.
- Geo-Coordinates and Statistics: The outcome of the study includes masks with geo-coordinates for integration into digital maps, and a dataframe containing comprehensive statistics about the damage in different zones.
Impact and Applications: This project represents a significant step in the application of AI to environmental studies, potentially aiding in conservation efforts, policy-making, and future forest management planning. By transforming raw data into actionable insights, the study illustrates the power of technology to address real-world challenges.
Acknowledgements: Special thanks to HackatonExpert Group and IPT faculty of Kyiv Polytechnic University for their invaluable support and collaboration.
