ESCNet: An End-to-End Superpixel-Enhanced Change Detection Network for Very-High-Resolution Remote Sensing Images

Area of Science:

• Earth Observation
• Computer Vision
• Remote Sensing

Background:

• Change detection (CD) is crucial in Earth observation, with recent advancements providing very-high-resolution (VHR) multispectral imagery.
• Precisely identifying changed areas in VHR images remains a significant challenge despite data enrichment.

Purpose of the Study:

• To propose an end-to-end superpixel-enhanced change detection network (ESCNet) for VHR images.
• To improve the accuracy of change detection by combining deep learning with superpixel segmentation.

Main Methods:

• Developed ESCNet, integrating differentiable superpixel segmentation and a deep convolutional neural network (DCNN).
• Utilized two weight-sharing superpixel sampling networks (SSNs) for feature extraction and superpixel segmentation of bitemporal image pairs.
• Employed a UNet-based Siamese neural network for difference mining, a novel superpixelation module for noise reduction, an adaptive superpixel merging (ASM) module, and a pixel-level refinement module.

Main Results:

• ESCNet effectively reduces latent noise in pixel-level feature maps while preserving edges.
• The adaptive superpixel merging (ASM) module compensates for superpixel number dependence and is fully differentiable.
• Experiments on public datasets demonstrated ESCNet's superiority over traditional and state-of-the-art deep learning-based CD (DLCD) methods.

Conclusions:

• ESCNet offers a robust and accurate solution for change detection in VHR imagery.
• The integration of superpixel techniques significantly enhances the performance of deep learning models in CD tasks.
• The proposed method advances the field of remote sensing by providing a more precise tool for identifying changes in Earth observation data.