Related Experiment Video
Updated: Aug 14, 2026
12:08
From Voxels to Knowledge: A Practical Guide to the Segmentation of Complex Electron Microscopy 3D-Data
Published on: August 13, 2014
Foundation Models for Remote Sensing Semantic Segmentation: A Review of Architectures, Adaptations, and Prospects
Ming Deng1, Yongyi Chen1, Guanghai Ding1
1College of Computer Science and Engineering, Guilin University of Technology, Guilin 541006, China.
Sensors (Basel, Switzerland)
|August 13, 2026
Summary
Foundation models are revolutionizing remote sensing image segmentation, moving beyond task-specific approaches. This review covers evolving datasets, new model architectures, and adaptation strategies for Earth observation tasks.
Area of Science:
- Earth Observation
- Computer Vision
- Geospatial AI
Background:
- Remote sensing image segmentation is crucial for Earth observation.
- Datasets are growing in scale, diversity, and complexity.
- The field is shifting towards foundation models.
Purpose of the Study:
- To review recent advancements in remote sensing image segmentation using foundation models.
- To analyze dataset evolution, model architectures, and adaptation strategies.
- To identify core challenges and future research directions.
Main Methods:
- Review of Transformer-based and Mamba-based architectures.
- Analysis of prompt-driven frameworks like Segment Anything Model (SAM).
- Examination of self-supervised and multimodal pre-training techniques.
Main Results:
- Foundation models, including Transformers and SSMs, are reshaping segmentation.
- Prompt engineering, parameter-efficient fine-tuning, and few-shot learning are key adaptation strategies.
- Challenges include balancing generality with specificity and handling sensor heterogeneity.
Conclusions:
- Foundation models offer powerful new paradigms for remote sensing image segmentation.
- Addressing challenges in adaptation and evaluation is critical for future progress.
- Future work should focus on remote-sensing-native pre-training and unified geospatial models.
