Segmentation models perform pixel-wise classification by assigning each pixel in an image to a specific class, enabling the identification of distinct objects or regions such as buildings, roads, water bodies, or crop fields. These models are widely used in domains like remote sensing, urban planning, and environmental monitoring, especially when applied to satellite and aerial imagery.

Traditionally, segmentation models require training from scratch using large, annotated datasets labelled with the desired object classes—a process that is both time-consuming and resource-intensive. Meta's Segment Anything Model (SAM) addresses this challenge by introducing a general-purpose, zero-shot segmentation model capable of identifying a wide range of objects without the need for additional training.

Trained on the Segment Anything 1-Billion mask dataset (SA-1B)—which includes over 11 million images and 1 billion segmentation masks—SAM is designed to generalize across domains and accurately detect object boundaries, even for objects it has never explicitly encountered. This makes SAM a powerful tool for extracting object masks from virtually any image, streamlining feature extraction workflows, and accelerating analysis across a variety of applications.