Error-Tolerant Deep Learning for Remote Sensing Image Scene Classification.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Remote sensing image scene classification (RSSC) is crucial for various applications.
• Deep convolutional neural networks (CNNs) excel in RSSC but require extensive labeled data.
• Acquiring accurate labels for large-scale remote sensing datasets is labor-intensive and prone to noise.

Purpose of the Study:

• To develop an effective error-tolerant deep learning approach for RSSC.
• To mitigate the negative impact of noisy labels in remote sensing datasets.
• To improve the robustness of CNN-based classification models in the presence of label errors.

Main Methods:

• Proposed a novel RSSC-oriented error-tolerant deep learning (RSSC-ETDL) approach.
• Implemented an iterative scheme alternating between learning multiview CNNs and correcting error labels.
• Introduced an adaptive multifeature collaborative representation classifier (AMF-CRC) for label correction.

Main Results:

• The RSSC-ETDL method demonstrated superior performance in handling noisy labels.
• Experiments were conducted on datasets with both simulated and real-world noisy labels.
• The proposed AMF-CRC effectively combined multiple CNN features for accurate label correction.

Conclusions:

• The RSSC-ETDL approach offers a robust solution for RSSC with noisy labels.
• The iterative learning and label correction strategy significantly enhances classification accuracy.
• The study provides a valuable method for practical remote sensing data analysis.