UNet-Like Remote Sensing Change Detection: A review of current models and research directions
Chen Wu, Liangpei Zhang, Bo Du, Hongruixuan Chen, Jingxuan Wang, Huan Zhong
Abstract
Recently, deep learning (DL) models have become the main focus for the remote sensing change detection tasks. Numerous publications on supervised and unsupervised DL-based change detection methods have been addressed. The end-to-end fully convolutional network has rapidly developed due to the release of more public datasets with labeled changes. Moreover, UNet is the most widely used basic structure for supervised DL-based models. Thus, this paper provides a novel categorical DL-based model review and systematically discusses the current UNet-like change detection methods. First, we divide the UNet-like model into seven basic blocks: encoder structure, symmetry, encoder module, feature to decoder, skip connection, data fusion, and loss function. Subsequently, we summarize the current UNet-like change detection publications and review various basic settings experiments. This review aims at providing a systematic overview of current DL-based change detection methods, offering insights into novel UNet-like models and highlighting the potential for future research.