Fuzziness-based active learning framework to enhance hyperspectral image classification performance for...

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Hyperspectral image classification demands extensive training data, which is often expensive and time-consuming to acquire.
• Limited training samples typically lead to significant generalization errors in classification models.

Purpose of the Study:

• To introduce a novel fuzziness-based active learning framework (FALF) for enhancing hyperspectral image classification with limited training data.
• To improve the generalization performance of both discriminative (e.g., SVM) and generative (e.g., KNN) classifiers.

Main Methods:

• FALF estimates class boundaries and calculates fuzziness-based distances for each sample to these boundaries.
• Samples closest to boundaries with higher fuzziness are selected as optimal training candidates.
• The framework is evaluated with Support Vector Machines (SVM) and K-Nearest Neighbors (KNN) classifiers.

Main Results:

• Classifiers trained using FALF demonstrated higher classification accuracy compared to random sample selection.
• The proposed method resulted in reduced processing times with limited training data.
• Experimental results on three public datasets confirm the effectiveness of FALF.

Conclusions:

• The fuzziness-based active learning framework (FALF) effectively addresses the challenge of limited training samples in hyperspectral image classification.
• FALF enhances classifier performance, achieving state-of-the-art results while being computationally efficient.