Cross-Scene Joint Classification of Multisource Data With Multilevel Domain Adaption Network.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Domain adaptation (DA) is crucial for applying models across different data distributions.
• Existing DA methods primarily address single-source, single-target scenarios.
• Integrating DA with multisource (MS) data collaboration presents significant challenges.

Purpose of the Study:

• To propose a novel multilevel domain adaptation network (MDA-NET).
• To enhance cross-scene (CS) classification by leveraging multisource hyperspectral image (HSI) and light detection and ranging (LiDAR) data.
• To promote information collaboration and improve classification accuracy in heterogeneous data environments.

Main Methods:

• Developed a multilevel DA network (MDA-NET) architecture.
• Incorporated modality-related adapters to process HSI and LiDAR data.
• Utilized a mutual-aid classifier to aggregate discriminative information across modalities.

Main Results:

• The MDA-NET demonstrated superior performance compared to state-of-the-art DA approaches.
• Experiments were conducted on two cross-domain datasets, validating the method's effectiveness.
• The proposed framework successfully boosted CS classification performance.

Conclusions:

• MDA-NET offers an effective solution for multisource domain adaptation in cross-scene classification.
• The method facilitates better information collaboration between HSI and LiDAR data.
• This work advances the capabilities of DA for complex remote sensing applications.