Robustness Evaluation of Hyperspectral Image Classification Models Using Degradation-Aware Metrics

Hyperspectral image classification studies are generally evaluated on clean benchmark datasets and often report high-accuracy values. However, because real sensor data can be affected by different degradations, such as Gaussian, impulse, stripe, dead-line, and mixed noise, clean-data performance alone may be insufficient to explain the reliability of models under practical conditions. In this study, SVM, k-NN, 1D-CNN, 2D-CNN, 3D-CNN, Hybrid CNN, Dense CNN, and a transformer-based model were evaluated on the Indian Pines, Pavia University, and Salinas datasets under controlled noise scenarios. The overall accuracy (OA), average accuracy (AA), kappa coefficient, and macro-F1 score were used as standard performance metrics. In addition, two robustness-oriented metrics, Relative Degradation Rate (RDR) and Robustness Score (RS), were used to measure noise-induced performance degradation more explicitly. The findings show that model rankings obtained on clean data may change under noisy conditions depending on the dataset and noise type. Performance degradation was more evident for some models, especially under mixed-severity and high-severity noise scenarios. Overall, the results indicate that hyperspectral classification models should be evaluated not only based on clean-data accuracy but also in terms of their robustness against different types of sensor noise.