Unsupervised Change Detection in Multitemporal VHR Images Based on Deep Kernel PCA Convolutional Mapping Network.

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Computer Vision

Background:

• Very-high-resolution (VHR) satellite imagery is crucial for change detection (CD).
• Deep learning (DL) methods show promise for CD but typically require extensive annotated training data.
• Existing DL models for CD often struggle with the need for labeled samples.

Purpose of the Study:

• To propose a novel unsupervised deep learning model for change detection in multitemporal VHR images.
• To develop a method that eliminates the need for annotated training samples in CD.
• To enhance the accuracy and robustness of change detection in VHR imagery.

Main Methods:

• Kernel Principal Component Analysis (KPCA) convolution for feature extraction.
• A deep siamese network (KPCA-MNet) with weight-shared KPCA convolutional layers.
• Mapping feature differences to a 2-D polar domain for segmentation and clustering.

Main Results:

• The proposed KPCA-MNet achieves effective change detection without labeled data.
• Experimental results on binary and multiclass CD datasets demonstrate the model's validity.
• The method shows robustness and potential for practical applications in VHR image analysis.

Conclusions:

• The unsupervised KPCA-MNet offers a viable alternative to supervised DL methods for VHR image change detection.
• The novel feature extraction and mapping techniques are effective for identifying spatial-spectral changes.
• This approach advances unsupervised change detection capabilities in Earth observation.