AI-Based Road and Tree Segmentation Service for Map Services

Project Requirement

AI-Based Road and Tree Segmentation Service for Map Services

A client in the map services industry had a service to automatically identify and extract roads and trees from their Geospatial and Aerial Survey (GAS) data. The process involved human operators going through thousands of GeoSpatial data and manually highlighting roads and tree crown segments. This process was manual, making it slow, expensive, and difficult to scale. Client wanted to scale up rapidly across regions and this was becoming a bottleneck. They wanted an automated solution that could accurately segment roads and tree crown, saving time and reducing costs while making their service more scalable.

Our Approach

We started with open-source geospatial segmentation datasets but found them insufficient for our needs. So, we built a hybrid dataset combining public data, manually segmented images, and GAN-generated samples. Using data augmentation and a custom CNN model, we trained on AWS SageMaker. The model was deployed as a scalable microservice using SageMaker and EKS, and retrained weekly with new human-labeled data to improve accuracy and performance.

Outcome and Impact

The model has reached an optimised level where it was producing highly accurate segmentation maps. Our micro services application completely automated the client's workflow, and completely eliminated manual generation. This allowed the client to expand their services globally across three regions in a year. They were able to scale it for different terrains and very high spatial resolutions (till 0.1m).