| Remote sensing image change detection is the process of detecting,identifying and quantifying surface feature changes through the analysis of multi-temporal remote sensing data.As an important subfield of remote sensing image change detection,road change detection can be widely applied to various scenarios,including road condition monitoring,urban planning and development and disaster emergency response and rescue.In current road change detection tasks,due to the difficulty of unifying road change definitions and annotating road change areas,pixel-level road change datasets are virtually nonexistent.Moreover,existing remote sensing image change detection algorithms use convolutional neural networks as feature extractors.However,due to limitations in the number of parameters and the design of local perception in convolutional layers,convolutional neural networks have a limited receptive field and cannot effectively capture more comprehensive and accurate high-level semantic information of road features.To establish a pipline for road change detection,including the construction of pixel-level road change datasets and the design and optimization of road change detection algorithms,this paper carried out the following research:1.For the construction of pixel-level road change datasets,this paper defines road changes as road damage changes and road addition/deletion changes.Road change areas are annotated from two perspectives to construct a pixel-level road change dataset.2.For the design of road change detection algorithms,this paper employs Transformer networks as feature extractors,using their global receptive field and excellent feature learning capabilities to obtain more comprehensive and accurate high-level semantic features of road information,thus reducing road pseudo-changes caused by shadows and occlusions.3.For the optimization of road change detection algorithms,from a data perspective,this paper uses data preprocessing to achieve style transformation between dual-temporal remote sensing images,eliminating the differences in spectral and spatial background features in the unchanged areas of the dual-temporal images and reducing the interference of irrelevant feature changes on target feature changes.From a model perspective,this paper constructs an unlabeled remote sensing image pretraining dataset and uses self-supervised learning methods to obtain pretraining weights based on the remote sensing image dataset,eliminating the domain differences between natural and remote sensing images that introduce systematic errors in downstream road change detection tasks.In summary,the comprehensive road change detection solution based on remote sensing images proposed in this paper has significant theoretical and practical application value.It can provide better support for various scenarios,such as road condition monitoring,urban planning and development,and disaster emergency response and rescue. |
