Distribution-Independent Domain Generalization for Multisource Remote Sensing Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Multisource remote sensing data enables comprehensive Earth observation.
• Convolutional Neural Networks (CNNs) integrate feature extraction and classification but struggle with domain generalization due to distribution assumptions.
• Limited access to testing data hinders the performance of traditional CNNs in cross-domain classification tasks.

Purpose of the Study:

• To develop a novel network, the feature-distribution-independent network (FDINet), for robust multisource remote sensing cross-domain classification.
• To address the generalization limitations of CNNs when faced with unseen target domains without requiring feature alignment or decoupling.
• To enhance the collaboration capabilities between different data sources and domains.

Main Methods:

• Designed a baseline network for extracting multisource cross-domain features, utilizing shallow weight-sharing networks to capture common line and texture features.
• Incorporated modality prediction probability to quantify source-target domain similarity, thereby boosting cross-domain collaboration.
• Developed a sharpness-aware feature discriminating (SAFD) strategy for model optimization, minimizing local optima sharpness to improve generalization.
• Introduced discrimination constraints to balance feature discrimination and generalization, mitigating gradient conflicts.

Main Results:

• The proposed FDINet demonstrated superior performance in multisource remote sensing cross-domain classification compared to existing methods.
• Experiments on two datasets validated the effectiveness of FDINet in both quantitative and qualitative analyses.
• The SAFD strategy successfully improved model generalization ability without sacrificing feature discrimination.

Conclusions:

• FDINet offers a powerful solution for cross-domain classification in multisource remote sensing, overcoming the limitations of traditional CNNs.
• The feature-distribution-independent approach and sharpness-aware optimization are key to achieving robust generalization.
• This work advances the application of deep learning in remote sensing for scenarios with limited or no access to target domain data.