Extreme Self-Paced Learning Machine for On-Orbit SAR Images Change Detection

Shuyuan Yang, Zhi Liu, Quanwei Gao, Yuteng Gao, Zhixi FengIEEE Access 2019 · OpenReview

Introduction

The rapid advancement of Earth observation satellites has created a surge in Synthetic Aperture Radar (SAR) data. Traditionally, this data is transmitted to ground stations for change detection—identifying surface alterations like flooding or deforestation. However, the massive volume of data often exceeds transmission bandwidth, creating a significant bottleneck. On-orbit processing—performing analysis directly on the satellite—offers a solution but faces two major hurdles: limited hardware resources (power and memory) and the lack of human-labeled training data in space.

To address these challenges, this paper introduces the Extreme Self-paced Learning Machine (ESLM). ESLM is a novel framework designed for autonomous, high-speed, and accurate change detection on-orbit, eliminating the need for ground-based intervention.

Methodology

The ESLM framework operates through a three-stage pipeline designed to maximize accuracy while minimizing computational complexity.

Overview of the ESLM framework for on-orbit SAR change detection, illustrating the transition from raw SAR images to final detected changes through self-paced learning.

1. Multistage Clustering (The "Teacher")

The process begins by segmenting SAR images into "super-pixels" to reduce noise sensitivity. Using Affinity Propagation Super-pixel Clustering (APSC), the system automatically categorizes these into three groups: strictly changed, strictly unchanged, and fuzzy (uncertain) pixels. This stage provides an initial "pseudo-labeled" training set autonomously.

2. Affinity Regularized Semi-supervised Extreme Learning Machine (AR-SELM)

The core classifier utilizes an Extreme Learning Machine (ELM). ELMs are ideal for satellite hardware because their internal weights are assigned randomly rather than through slow, iterative processes, leading to extremely fast training. The authors incorporate a manifold regularizer (AR-SELM) to allow the model to learn from both the labeled and the unlabeled (fuzzy) pixels simultaneously.

3. Self-paced Learning (The "Curriculum")

ESLM adopts a "simple-to-complex" learning strategy. It initially trains on the most confident pixels (the "easy" examples) and gradually incorporates more difficult, fuzzy pixels into the training set. This incremental approach prevents the model from being misled by the "speckle noise" common in radar imagery, leading to more robust decision boundaries.

Experimental Evaluation

The researchers evaluated ESLM on five benchmark SAR datasets, including flood detection in Bern and Ottawa, and estuary monitoring in the Yellow River.

Key Results

  • Accuracy: On the Bern dataset, ESLM achieved a Kappa Coefficient of 0.908, outperforming several state-of-the-art methods.
  • Efficiency: ESLM demonstrated a significant speed advantage. While deep learning approaches like Local Restricted Convolutional Networks (LRCN) required nearly 250 seconds for the Bern dataset, ESLM completed the task in just 21.6 seconds—a 10x speedup.
  • Robustness: Qualitative analysis showed that ESLM produces cleaner change maps with fewer false alarms compared to traditional dictionary learning or autoencoder methods.

Conclusion

The Extreme Self-paced Learning Machine (ESLM) provides a compelling solution for on-orbit SAR change detection. By mimicking human learning patterns and leveraging the speed of ELMs, it achieves high accuracy with low computational overhead. While challenges remain regarding extremely heavy speckle noise in single-look images, ESLM represents a significant step toward autonomous, real-time satellite intelligence. Future work will focus on integrating advanced noise-reduction techniques and exploring deeper self-paced architectures.

Summary updated

Abstract

With the rapid development of earth observation satellites, on-orbit data processing is becoming more and more desirable. In this paper, a new on-orbit change detection method for Synthetic Aperture Radar (SAR) images, is proposed via an Extreme Self-paced Learning Machine (ESLM). First, a reflectivity-spatial affinity is defined to measure the similarity between two segmented super-pixels, to identify the initial three groups of pixels: strictly changed, strictly unchanged and fuzzy pixels. Then a new extreme self-paced learning machine is developed, by gradually selecting the most confident changed pixels and predicting the changed pixels in an incremental pattern. Moreover, both the labeled and unlabeled samples are explored to realize semi-supervised classification. Different with the available methods, ESLM works in a self-paced learning pattern and achieves accurate detection, for it can automatically choose the training samples and explore unlabeled samples to enhance the online prediction of changes. Therefore, ESLM has the characteristics of accurate and robust detection, parameter free, low-complexity and rapid implementation, which is very suitable for on-orbit processing. Some experiments are taken on five real benchmark datasets, and the results verify the effectiveness of ESLM.
BibTeX
@inproceedings{yang2019extreme,
  author = {Shuyuan Yang and Zhi Liu and Quanwei Gao and Yuteng Gao and Zhixi Feng},
  title = {{Extreme Self-Paced Learning Machine for On-Orbit SAR Images Change Detection}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=n05fCCpEuy},
}
Figure