Multi-Label Noise Robust Collaborative Learning for Remote Sensing Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Accurate multi-label classification (MLC) of remote sensing (RS) images is crucial.
• Convolutional neural network (CNN) based MLC methods require extensive, accurately labeled data, which is costly and time-consuming to acquire.
• Publicly available thematic products offer a cost-effective annotation source but often contain noisy labels, hindering model performance.

Purpose of the Study:

• To develop a novel method for robust multi-label classification in remote sensing images.
• To address the challenges posed by multi-label noise (incorrect and missing labels) in training data.
• To improve the performance of CNN models for RS image MLC despite the presence of noisy labels.

Main Methods:

• Proposes a robust multi-label noise-robust collaborative learning (RCML) method.
• RCML utilizes three modules: discrepancy, group lasso, and swap, to identify, rank, and exclude noisy labels.
• The method does not make prior assumptions about the noise distribution in the training data.

Main Results:

• Demonstrates the robustness of the RCML method under extreme multi-label noise rates.
• Experimental results on two multi-label RS image archives validate the effectiveness of the proposed approach.
• The developed RCML method alleviates the negative effects of multi-label noise during CNN model training.

Conclusions:

• The proposed RCML method offers a robust solution for multi-label classification of remote sensing images with noisy labels.
• RCML enhances the reliability of training data by effectively handling label noise without specific noise distribution assumptions.
• The publicly available code facilitates further research and application of noise-robust learning in remote sensing.