| Title |
A hybrid dynamic graph neural network-XGBoost framework for hyperspectral urban scene classification |
| Authors |
Albarakati, Hussain Mobarak ; Hamza, Ameer ; Alabdullah, Bayan ; Alasiry, Areej ; Marzougui, Mehrez ; Yang, Zepa |
| DOI |
10.1109/JSTARS.2026.3716175 |
| Full Text |
|
| Is Part of |
IEEE Journal of selected topics in applied earth observations and remote sensing.. Piscataway, NJ : IEEE. 2026, Early access, p. 1-28.. ISSN 1939-1404. eISSN 2151-1535 |
| Keywords [eng] |
dynamic graph neural network ; graph-based representation learning ; Hyperspectral image classification ; imbalanced data classification ; non-local feature relations ; spectral-spatial analysis ; urban scene classification ; XGBoost |
| Abstract [eng] |
High spectral dimensionality, non-local class connections, and unbalanced class distributions can all restrict the efficacy of fixed-neighborhood classifiers, making hyperspectral image classification difficult. In order to build discriminative representations while enhancing robustness to class imbalance, this study suggests a hybrid Dynamic Graph Neural Network XGBoost architecture for hyperspectral urban scene classification. The suggested method aggregates relevant non-local spectral relationships beyond strict local kernels by first encoding each labelled pixel using a dynamic graph neural network that builds adaptive connections in feature space. After that, the graph-aware embedding are sent to an XGBoost classifier, which improves class-wise separability and refines the decision boundaries. To optimise the boosting step and the graph encoder, a thorough grid search was carried out. While retaining nearly comparable validation performance (Overala Accuracy = 93.17%, macro-F1 = 0.9196), the best performing configuration obtained test overall accuracy of 93.01%, average accuracy of 91.15%, Cohen's kappa of 0.9072, and macro-F1 of 0.9167. While the majority of residual errors were focused among spectrally identical urban materials, class-wise examination also showed notably strong discriminating for spectrally distinctive and minority categories. These findings imply that the suggested hybrid approach successfully blends gradient boosted classification with adaptive graph-based representation learning, providing a viable substitute for traditional hyperspectral classifiers for intricate urban land-cover research. |
| Published |
Piscataway, NJ : IEEE |
| Type |
Journal article |
| Language |
English |
| Publication date |
2026 |
| CC license |
|
