Superpixel segmentation integrated feature subset selection for wetland classification over Yellow River Delta.

Area of Science:

• Ecology
• Remote Sensing
• Machine Learning

Background:

• Wetlands are critical ecosystems facing significant threats.
• The Yellow River Delta wetlands exhibit complex land cover due to riverine and oceanic influences.
• Accurate wetland classification is essential for conservation and management.

Purpose of the Study:

• To develop an effective method for distinguishing wetland types in the Yellow River Delta.
• To improve upon existing pixel-based classification limitations.
• To enhance the accuracy of wetland classification in dynamic environments.

Main Methods:

• Object-oriented image analysis combined with machine learning.
• Superpixel segmentation using the watershed algorithm with H-minima labeling.
• Recursive feature elimination cross-validation for optimal spectral index selection.
• Random Forest classifier integrating segmentation and feature selection.

Main Results:

• The proposed method significantly improved wetland classification accuracy.
• Achieved an overall accuracy of 91.74% and a kappa coefficient of 0.9078.
• Outperformed classical pixel-based machine learning methods, including pixel-oriented Random Forest.

Conclusions:

• The object-oriented, feature-preference machine learning approach is highly effective for Yellow River Delta wetland classification.
• This method overcomes limitations of pixel-based approaches, such as the 'pretzel phenomenon'.
• The study demonstrates a substantial improvement in classifying complex and dynamic wetland environments.