Sentinel 2 Bands + Deep Learning

Question

I am trying to classify 9 different classes of Land use deploying Resnet Architecture. Sentinel images come with 13 bands and in RS we normally combine different bands for different classes for example for agriculture color bands (11,8,3) are composites for agriculture, and so forth with RGB, True Color, NDVI.

I have tried combining different bands and so far accuracy is 0.83 with RGB. I tried SWIR and True Color combinations and both have lower scores.

In addition, I tried to have different and appropriate bands for each class, for example, Residential(Bands 4,3,2), River(8,11,4), and Crops(11,8,2). Well, this has improved the accuracy of the model up to 0.93.

My question is twofold:

1. Is this really an acceptable approach in Deep Learning practice? All the literature I have seen so far are using the same bands for each class.

2. If this approach is ok, then can I use different band combinations for the same class as data augmentation? this will increase data size and hence improve the accuracy.

Answer 1

1. 'acceptable approach' is a bit of a complicated issue, when it comes to deep learning, as the field is very result oriented in general. If something improves results, it is generally used, even if it is dubious.

2. Boot-strapping is also widely done and seen as a positive, at least in so far as I have seen.

My main concern is that you appear not be to using all the bands at the same time. As far as I remember (and this may be wrong), Resnet does support more than three bands, so you should be able to use all 13 bands. This does increase the size of the search space, but it shouldn't be too bad.

Answer 2

Create a stack with this bands :

B02_10m.jp2",_B03_10m.jp2","_B04_10m.jp2","_B05_20m.jp2","_B06_20m.jp2","_B07_20m.jp2", _B08_10m.jp2","_B8A_20m.jp2","_B11_20m.jp2" ,"_B12_20m.jp2"

and add in the stack a vegetation index.

Run the model.