SemanticSegmentation

Semantic segmentation on aerial images (aka image classification) using a CNN-based UNet.

Implementation

• trains from scratch a CNN with a UNet architecture with 3 (or 4) lateral connections: https://arxiv.org/pdf/1505.04597.pdf
• valid-padding is used to avoid border issues
• segmentation maps for each image patch are stitched back together to create complete map

Data

• 9 cm spatial resolution aerial images (3-channel: NIR-Red-Green) of the town of Vaihingen from the ISPRS 2D Semantic Labeling Contest: http://www2.isprs.org/commissions/comm3/wg4/2d-sem-label-vaihingen.html
• corresponding normalized Digital Surface Model (nDSM) from: https://www.researchgate.net/publication/270104315_Normalized_DSM_-_heights_encoded_in_dm_-_see_report_for_details
• we use NIR, Red and nDSM (discarding Green), as Keras' ImageDataGenerator accepts only 3 channels

Results

• after a few epochs of training on a laptop GPU: OA: 83.33, Kappa: 0.778 (on separate test set)
• misses classes "cars" and "clutter/background" (small nr of samples): stratified sampling to be implemented to solve the issue

Ground Truth	Prediction