A Robust Dynamic Classifier Selection Approach for Hyperspectral Images with Imprecise Label Information.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Supervised hyperspectral image (HSI) classification requires accurate training labels, which are often difficult to obtain.
• Existing classification methods are sensitive to label errors, limiting their real-world applicability.
• The impact of erroneous labels on HSI classification uncertainty remains under-investigated.

Purpose of the Study:

• To develop a robust classification model that can handle erroneous labels in hyperspectral data.
• To statistically analyze the uncertainty introduced by label errors in HSI classification.
• To propose a novel robust dynamic classifier selection (R-DCS) model for improved classification performance.

Main Methods:

• Analysis of the effect of label errors on principal component probability distributions in HSIs.
• Development of the R-DCS model based on the theory of imprecise probabilities and classifier prediction robustness.
• Extraction of spectral and spatial features to build individual classifiers for dynamic selection within the R-DCS framework.

Main Results:

• The proposed R-DCS model effectively mitigates the negative impact of erroneous labels on HSI classification.
• Experimental results on benchmark datasets show the R-DCS model outperforms individual classifiers.
• The R-DCS model demonstrates enhanced robustness to label errors compared to conventional approaches.

Conclusions:

• The R-DCS model offers a robust solution for hyperspectral image classification with imperfect label information.
• The developed model provides a statistically grounded approach to managing uncertainty arising from label noise.
• This work advances the field by providing a practical and effective method for HSI classification in real-world scenarios with noisy data.