Submission for EY Data Science Challenge 2024
To develop a machine learning model to identify and detect “damaged” and “un-damaged” coastal infrastructure (residential and commercial buildings), which have been impacted by natural calamities such as hurricanes, cyclones, etc.
- Region: Puerto Rico
- Storm: Hurricane Maria in (2017)
- Pre Event Date - 29th Aug’17
- Post Event Date - 12th Oct’17
- Clouds appear as saturated (dark) regions in each color band. Pre-storm image has more clouds than post storm image.
- No cloud masking available, although there are clouds in the images.
- Inherent difficulty in differentiating resedential and commercial buildings.
- TIFF: Tag Image File Format. Lossless form of file compression (no image quality loss).
-LabelMe for Annotation
-YOLOv8 deep-learning model
Break up both pre and post hurrican images into grid. Convert grid tif to jpeg (for compatibility with LabelMe).
Compare pre images with post while manually annotating.
Given that pre-images, as the name suggests, were taken prior to the Hurricane, the buildings will be undamaged. This gives the annotator a reference for identifying damaged and undamaged images, and the model could potentially be trained on pre-images for undamaged residential and commerical buildings.