Classify pixel using deep learning

[sharmaakki]

Dear All

I have trained a Unet and Deeplab V3 model for pixel classification using deep learning but the in the output I am not getting a uniform classified raster. Instead of that class shapes have been shown in the output that too not for the whole image.

I am attaching model parameters and classified pixel screenshot for reference.

Hoping for a reply soon.

Thanks in advance.

[Anonymous User]

Hi @sharmaakki ,

It seems that your model is still training you can continue training it for some more time. You can supply the previous best trained model as pretrained model and further train it. For best results you should continue training it till the time see that loss is not going down over some epochs and accuracy is not improving yo

Thanks,

Sandeep

[PhamMinhThang]

Taking a look at your text file, you only trained your model for 30 epochs, so maybe your model did not reach the optimal performance. Try to set epochs at 100, or 200, but choose Stop when model stops improving = STOP_TRAINING to prevent overfitting.

You used multi classes for training deep learning model. So considering collect more samples for each class to avoid the imbalance.

Try different backbone, maybe Resnet-34 is not suitable for your task.

OMG, this post was years ago.