Multimodal remote sensing benchmark datasets for land cover classification with a shared and specific feature...

Area of Science:

• Geoscience
• Remote Sensing
• Data Science

Background:

• Increasing availability of large-scale, open remote sensing (RS) data from diverse sensors.
• Challenges in effectively integrating complementary information from heterogeneous RS data modalities due to differences in sensors, resolutions, and content.
• Need for consistent, compact, accurate, and discriminative representations for multimodal RS data analysis.

Purpose of the Study:

• To propose a novel model for effective multimodal remote sensing data fusion.
• To address the challenge of embedding complementary information from heterogeneous data sources into a unified representation.
• To enhance land cover classification accuracy using multimodal RS data.

Main Methods:

• Development of a shared and specific feature learning (S2FL) model.
• Decomposition of multimodal RS data into modality-shared and modality-specific components.
• Utilizing three benchmark datasets (Houston2013, Berlin, Augsburg) for model evaluation.

Main Results:

• The proposed S2FL model demonstrates superior performance in land cover classification compared to state-of-the-art baselines.
• Effective information blending across heterogeneous data sources is achieved.
• Experimental validation on diverse multimodal datasets confirms model advancement.

Conclusions:

• The S2FL model offers an effective approach for multimodal remote sensing data analysis and fusion.
• The model's ability to handle heterogeneous data sources significantly improves land cover classification.
• Released benchmark datasets and baseline codes facilitate further research in multimodal RS data processing.