fbpx Solar startup - Omdena

ML in Solar: Identifying rooftops on low-resolution images

Location: Remote

Building an ML model to identify rooftop edges in low-resolution pictures is a hard problem that has not been solved before. Solutions like Google Sunroof project work only on high-resolution images and are therefore not usable in the majority of the developing world.

This did not stop the Indian startup Savera to take on the challenge together with Omdena.

Problem: The low quality of satellite images in India makes it impossible to automatically identify rooftop edges for higher solar panel adoption.

The project requirements:

The data input: Low-resolution pictures of rooftops
Desired model outcome: Build a machine learning model to identify rooftop areas on low-resolution satellite images.
Data availability: The raw image data is available but rooftops edges need to be tagged manually.

Data availability and requirements: Raw image data is available but rooftops edges need to be tagged manually.

The Project Community:

40-50 talented students and juniors, living in remote parts of India who never met, working collaboratively with two mentors and one community manager to generate and prepare the data while trying different models through our Collaborative AI model.

Result:

The community tried five different models including deep learning models (U-Net and Masked R-CNN) and over 1000s of datasets were tagged by community members. The best results were found by using the Masked R-CNN model trained on coco dataset with transfer learning. The quality of the model has been the highest compared to other available solutions and has been featured in leading Data Science Blogs. More details can be found here and here.

Want to work on similar projects?