Difficulty in Training Deep Learning Model

[IIRSIT]

Hello there!

I tried to use deep learning model to train through U-Net for the classification of flood and permanent waterbody, I tried on varied epochs as well; ranging from 20, 40 and 60 yet the accuracy of waterbody is not that good as compared to flood. 

I am facing difficulty in getting good results and accuracy for classification of flood.  

I am referring to this notebook: https://developers.arcgis.com/python/samples/flood-inundation-mapping-using-sar-data-and-deep-learni... for help. 

How can I get better results?

Attaching a screenshot of the model metrics for reference. 

with 40 epochs: 

with 60 epochs:

Kindly help me out.. 

Aishwarya M

[ShivaniPathak]

Hi Aishwarya,

Which sensor are you working? how many image chips you have in your training data?

[IIRSIT]

Hello! Thanks for your prompt response.

I am working on Sentinel 1- C band SAR data for VV nd VH both polarizations. I can share the training data with you for your reference. 

Sharing my mail ID to communicate further - aishwaryamakwana99@gmail.com

[ShivaniPathak]

Hi,

Please share your training data.

[ShivaniPathak]

As the training area is very small. Best approach for training the model will be by using both polarization bands VV & VH. A composite raster can be created which will have Band 1 = VV, Band 2 = VH and Band 3 = VH + VV. We are using VH + VV because literature suggests it works better for urban flooding. Use this composite as input raster. 

For increasing accuracy you can increase the training data area.

[aishwaryamakwana]

Hello! 

I was able to achieve good results and accuracy using this approach. Thanks for the solution.