Devices, methods, and program storage devices for training and leveraging machine learning (ML) models to use in image registration, especially on unaligned multispectral images, are disclosed, comprising: obtaining aligned multispectral image data; generating a first plurality of feature descriptors for features identified in the aligned multispectral image data; generating a training set of feature descriptor pairs based on the first plurality of feature descriptors; and training a ML model based on the training set of feature descriptor pairs, wherein the trained ML model is configured to determine matches between features in unaligned multispectral image data. The techniques may then: obtain unaligned multispectral image data; generate a second plurality of feature descriptors for features identified in the unaligned multispectral image data; and use the trained ML model to determine matches between features in the second plurality of feature descriptors, which matches may be used in performing image registration and/or fusion operations.