Asymmetric Feature Fusion Network for Hyperspectral and SAR Image Classification

Area of Science:

• Remote Sensing
• Earth Observation
• Computer Vision

Background:

• Joint classification of multisource remote sensing data is crucial for Earth observation but faces challenges due to differing imaging mechanisms and imbalanced information.
• Integrating complementary information from diverse data sources for accurate interpretation remains a significant hurdle in remote sensing.

Purpose of the Study:

• To propose an effective method for joint classification using multisource remote sensing data.
• To address the difficulties in fusing heterogeneous data by developing an asymmetric feature fusion network (AsyFFNet).

Main Methods:

• Developed the asymmetric feature fusion network (AsyFFNet) utilizing weight-share residual blocks for feature extraction with separate batch normalization (BN) layers.
• Implemented a channel redundancy determination mechanism within BN layers, where scaling factors identify and replace redundant channels.
• Incorporated a sparse constraint on BN scaling factors to eliminate unnecessary channels and enhance model generalization, alongside a feature calibration module to leverage spatial dependencies.

Main Results:

• The proposed AsyFFNet demonstrated superior performance in joint classification tasks compared to existing methods.
• Experimental results on three distinct datasets validated the effectiveness of the asymmetric feature fusion approach.

Conclusions:

• The AsyFFNet provides a robust solution for overcoming the challenges of multisource remote sensing data fusion.
• The method significantly enhances classification accuracy and generalization capability in Earth observation applications.