研究目的
Exploring the usage of convolutional neural networks for urban change detection using multispectral images from the Copernicus Sentinel-2 program.
研究成果
The paper presents the OSCD dataset for urban change detection and two CNN approaches that achieve excellent test performances. Future work includes enlarging the dataset and experimenting with fully convolutional networks for automatic label generation.
研究不足
The dataset's images are of relatively low resolution, limiting the detection of small changes. The temporal distance between images is at most two and a half years, and the dataset contains significantly more pixels labelled as no change than change.
1：Experimental Design and Method Selection:
The study employs two CNN architectures, Siamese and Early Fusion, for change detection.
2：Sample Selection and Data Sources:
The OSCD dataset, composed of pairs of multispectral aerial images with manually annotated changes at pixel level, is used.
3：List of Experimental Equipment and Materials:
Sentinel-2 satellite images, Medusa toolbox for image cropping, GEFolki toolbox for image registration.
4：Experimental Procedures and Operational Workflow:
Training the proposed networks on the OSCD dataset, evaluating the impact of different numbers of spectral channels as inputs.
5：Data Analysis Methods:
Comparison of classification performance between the two architectures and different input channel configurations.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
