Browsing through the source code shows that it requires a data folder with 3 subfolders (images_before, images_after, labels). I presume the images should have the same name in each folder. The label images should use black (RGB 0,0,0) for No Change and white (RGB 255,255,255) for Change.
There is a link to the research paper in the help:
https://www.mdpi.com/2072-4292/12/10/1662/htm
From the research paper you can find links to a sample change detection training dataset on GitHub:
https://justchenhao.github.io/LEVIR/
If you download the files, just create the required subfolders listed above and copy the relevant images into each one.
I have managed to get it training in a notebook, but the results are very underwhelming so far. The sample images above are 1024x1024 with lots of No Change and I think that may be causing issues.
from arcgis.learn import ChangeDetector, prepare_data
data_path = r'<path_to_data_folder>'
data = prepare_data(data_path, batch_size=1, dataset_type='ChangeDetection')
model = ChangeDetector(data)
model.fit(10, lr=1e-4)
model.show_results()
model.save(r'<path_to_data_folder>',save_inference_file=True)
