Change Detection Dataset
As a common dataset for change detection, its image can be divided into three parts: the image before the change, the image after the change, and the label image showing the changed area. This dataset is characterized by significant seasonal differences between bi-temporal image pairs, which makes up for some of the deficiencies in existing datasets. The labels for this dataset include some irregular changes, such as the appearance and disappearance of cars; but do not include seasonal changes, such as changes in the ground surface caused by snowfall.
This dataset contains 11 pairs of multispectral images obtained by Google Earth, including 7 pairs of 4725 × 2200 pixel very high resolution (VHR) seasonal variation images and 4 pairs of 1900 × 1000 pixel images, Image resolutions vary from 3 cm to 100 cm per pixel. The authors of the dataset cropped and randomly rotated these images, resulting in 1600 bi-temporal image pairs, where each image was a 256 × 256 pixel 3-channel remote sensing image. The ratio of training set, validation set, and test set is 10:3:3.