Our client is a leading provider of commercial business insurance. Their innovation team intended to apply AI and advanced data science capabilities to conduct accurate risk assessments on properties by analyzing images captured by drones.
More specifically, the client wanted to:
- Identify vegetation regions from the normal and infrared aerial images
- Segment regions into dense/sparse vegetation and identify buildings, roads, cars, parking lots, etc.
- Limited number of images available to train AI models
- Selecting the right methodologies for segmentation given the limited data available for analysis
- Deployed leading data augmentation techniques to artificially create training images by introducing deliberate variations in brightness, contrast, and orientation
- Leveraged Normalized Difference Vegetation Index (NDVI) to identify vegetation condition in an infrared image at a pixel level
- Used U-NET deep learning model for multi-attribute segmentation – manmade vs. vegetation, parking lot vs. building vs. road
- The AI solution is being utilized by the client to:
- Ascertain property risk exposure from forest fires based on dense and sparse vegetation cover identified by the models
- Evaluate the condition of manmade structures such as parking lots and roads