Satellite Image Segmentation

Solution that we have achieved on GIS Tech Hack hackathon on building segmentation

Hackathon task

The task on that hackathon was to make a mask of buildings on a given image taken from satellite. The image was taken in Russia, Republic of Tatarstan, Kazan, Aviastroitelny district.

The organizers has provided us with servers for computing. These servers have GPU NVIDIA TESLA V100 with 32GB RAM on board

The input image size was 6528x7734 pixels with 3 channels (RGB).

Solution:

1. The original images and train masks were cropped to the image size 256x256 with 3 channels.
2. The architecture of neural network was similar to U-Net but with some modifications.
   • As encoder part we have taken pretrained ResNet50 convolution network on imagenet dataset
   • Decoder model have 5 upsampling layers. On each layer (exept the last) there is 2 convolution layers with batch normaliztion
3. You can download weights here

Results