An Impartial Semi-Supervised Learning Strategy for Imbalanced Classification on VHR Images.

Area of Science:

• Remote Sensing
• Machine Learning
• Geospatial Analysis

Background:

• Imbalanced learning is a significant challenge in remote sensing land-use/land-cover classification.
• This imbalance can decrease classification accuracy and cause minority classes to be overlooked.

Purpose of the Study:

• To propose an impartial semi-supervised learning strategy (ISS-XGB) for classifying imbalanced very high resolution (VHR) remote sensing imagery.
• To address the limitations of existing methods in handling imbalanced remote sensing data.

Main Methods:

• Developed an impartial semi-supervised learning strategy (ISS-XGB) utilizing extreme gradient boosting.
• Employed multi-group unlabeled data to mitigate training sample imbalance.
• Utilized gradient boosting-based regression with positive and unlabeled samples to simulate target classes.

Main Results:

• ISS-XGB demonstrated comparable and more stable performance than Random Forest, XGB, MLP, SVM, PU-BP, PU-SVM, and synthetic sample methods.
• Achieved high accuracy for minority classes without compromising overall performance, with an average overall accuracy of 85.92% under extreme imbalance.
• Experiments were validated across eight diverse study areas with varying imbalance levels.

Conclusions:

• ISS-XGB offers a robust solution for imbalanced classification problems in remote sensing.
• The proposed strategy shows significant potential for improving land-use and land-cover classification accuracy with imbalanced VHR imagery.