From Data to Knowledge: A Knowledge Graph-Guided Framework to Deep Learning for Hyperspectral Image Classification
Recent advances in deep learning have significantly improved hyperspectral image (HSI) classification. However, deep learning models for HSI classification typically rely on one-hot labels, which lack semantic information and fail to reflect relationships between land cover classes, leading to subop...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11029574/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Recent advances in deep learning have significantly improved hyperspectral image (HSI) classification. However, deep learning models for HSI classification typically rely on one-hot labels, which lack semantic information and fail to reflect relationships between land cover classes, leading to suboptimal generalization and interpretability. To overcome these problems, we propose a novel knowledge graph (KG)-guided HSI classification framework that bridges symbolic reasoning and connectionist learning. Unlike previous methods, our framework incorporates a KG to encode explicit symbolic knowledge, providing a structured definition of land cover classes. We employ KG embedding techniques to transform symbolic knowledge into continuous vector representations, seamlessly integrating structured semantics with deep learning models. Furthermore, we develop two KG integration approaches: a regression-based method and a classification-based method, demonstrating the complementary role of structured symbolic knowledge in enhancing connectionist learning. Extensive experiments on three representative HSI datasets show that integrating KG significantly improves the performance of a wide range of deep learning architectures in HSI classification, highlighting the broad applicability and robustness of the proposed framework across diverse deep learning architectures. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |