Machine Learning

Semantic Segmentation

A deep learning technique that classifies every pixel in an image into a predefined category. In Earth observation, it's used to create detailed land cover maps, delineate crop fields, detect buildings, and map water bodies at pixel-level precision.

Overview

Semantic segmentation assigns a class label to every pixel in an image, producing a dense classification map. In earth observation, this transforms satellite imagery into detailed thematic maps where every pixel is labeled as water, forest, urban, cropland, etc. It is the backbone of automated land cover mapping at continental and global scales.

How It Works

Models use encoder-decoder architectures. U-Net uses skip connections to fuse fine spatial details with abstract features. DeepLab introduced Atrous Spatial Pyramid Pooling for multi-scale context. SegFormer uses transformer self-attention for long-range spatial relationships. Training requires pixel-level labeled datasets. Evaluation uses mean Intersection over Union (mIoU).

Key Facts

  • U-Net, DeepLab, and SegFormer are the most widely used architectures.
  • Does not distinguish individual instances — two adjacent buildings are one "urban" region.
  • State-of-the-art achieves mIoU above 0.90 on standard remote sensing benchmarks.
  • Multi-temporal approaches using time series improve seasonal land cover classification.

Applications

Land Cover Mapping

Classifying every pixel into categories like forest, water, urban, cropland, and wetland.

Crop Type Classification

Identifying specific crop species from multispectral time series imagery.

Flood Mapping

Rapidly delineating flooded areas from SAR or optical imagery during disasters.

Limitations & Considerations

Requires large volumes of pixel-level labeled training data. Struggles with class boundaries in mixed-pixel areas. Performance degrades across different geographic regions without domain adaptation. High-resolution imagery demands significant computational resources.

History & Background

Evolved from pixel-based classifiers in the 1970s-80s. FCNs (2015) demonstrated end-to-end pixel classification. U-Net and DeepLab followed. SegFormer brought transformers in 2021. Foundation models like Prithvi now provide pre-trained encoders.

Analyze Semantic Segmentation data with LYRASENSE

Use our agentic notebook environment to work with satellite data and compute indices like Semantic Segmentation — no setup required.