Advancing Weakly-Supervised Change Detection in Satellite Images via Adversarial Class Prompting.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Weakly-Supervised Change Detection (WSCD) uses image-level labels for satellite image analysis, reducing annotation needs.
• WSCD methods often misclassify background variations as object changes due to limited supervision.
• Complex remote-sensing scenarios exacerbate the challenge of distinguishing object changes from background noise.

Purpose of the Study:

• To address the co-occurring noise problem in Weakly-Supervised Change Detection.
• To develop a method that accurately differentiates object changes from background variations.
• To enhance the performance of existing WSCD techniques without increasing inference costs.

Main Methods:

• Propose Adversarial Class Prompting (AdvCP) with two phases: Adversarial Prompt Mining and Adversarial Sample Rectification.
• Adversarial Prompt Mining uses incorrect labels to reveal background features misclassified as object changes.
• Adversarial Sample Rectification integrates these samples using an online global prototype for unbiased correction.

Main Results:

• AdvCP significantly improves performance across ConvNet, Transformer, and Segment Anything Model (SAM) baselines.
• Achieved substantial IoU improvements: 7.37% on WHU-CD, 7.46% on LEVIR-CD, and 6.56% on DSIFN-CD.
• Demonstrated generalizability to other multi-class weakly-supervised dense prediction tasks.

Conclusions:

• AdvCP effectively mitigates background noise in WSCD, enhancing accuracy.
• The method seamlessly integrates with existing WSCD approaches, offering practical benefits.
• AdvCP shows promise for improving change detection in various remote sensing applications.