I am currently working on a project that tackles crop type classification using machine learning and deep learning. Given the limited amount of satellite data that I have, transfer learning sounds like the next best option. I was wondering how can I find pre-trained models?


Pre-trained models generally rely on the pre-processing of the imagery being very consistent. As such, it is generally not feasible to do transfer learning with a remote sensing based classification model that was not trained by yourself, since you will rarely be able to do the exact same pre-processing.

All in all, I'd not recommend attempting transfer learning, using a model you did not train yourself.

  • 1
    This highly depends on which layers you decide to cut off and which ones you decide to keep... If you keep only low level features layers, it will certainly work better as starting from scratch, hence saving a large amount of time. May 19 '20 at 9:02
  • True - using pre-trained weights can be useful, but you still need good training data, and the pre-trained weights also need to be designed for the correct number of bands. May 19 '20 at 9:31
  • @Mikkel Lydholm Rasmussen Then does this mean that when looking for a pre-trained model I need to find one that was trained on similar features that I am planning to use?
    – Rim Sleimi
    May 19 '20 at 16:07
  • 1
    @RimSleimi - yes. When doing deep learning on remote sensing imagery, some efficiency can be gained by using pre-trained weights from stuff like ImageNet, but it also limits your options in terms of input data, as many pre-trained weights only work with 3-band natural images, like you would see in a normal camera. As such, you'd want to find a set of weights that fit your purpose and there aren't that many options out there.. And those that are may well be focused on something entirely else than what you want to map. May 20 '20 at 6:55
  • @MikkelLydholmRasmussen that what I have noticed, all the existing models are targeted towards RGB images and thus cannot be applied to remotely sensed images, especially that I am planning to use NDVI, NDWI, and all the spectral bands for crop type classification.
    – Rim Sleimi
    May 20 '20 at 13:58

You can use the pre-trained models (tensorflow) from a remote sensing dataset, here is the golden standard for this domain: http://bigearth.net/

  • My understanding is that bigearthnet is a benchmark dataset, not a model. Thus, I don't understand how this answer relates to the question asking about pre-trained models. Am I missing something?
    – Jeremy
    May 10 at 20:23
  • you can find "Deep Learning Models" on the page
    – rnd_nr_gen
    May 24 at 17:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.