What is semantic segmentation of satellite images and which models are used?

Semantic segmentation of satellite images is a task of labeling each pixel of the image with a class (e.g., water, forest, city, agricultural land). This allows for detailed analysis of the Earth's surface.

Deep neural networks are used to solve this, in particular:

• U-Net — a popular architecture for segmentation, works well with limited data.
• DeepLab (v3, v3+) — uses spatial pyramids and atrous convolutions to improve segmentation quality.
• SegNet — an encoder-decoder architecture, effective for high-resolution tasks.

An important aspect is data preprocessing (correction of atmospheric distortions, normalization) and the use of multispectral channels to improve classification accuracy.