State of the art model for classifying 256 crop categories built in Keras.
U-Net
https://github.com/ZihengSun/Ag-Net-Dataset
https://github.com/ZihengSun/Ag-Net-Dataset/blob/master/colormap.py
After training for 1000+ epochs, I was able to get training accuracy of about 91% and validation accuracy of 84%.
-
Use of skip connections.
-
Use of GlobalMaxPool2D instead of MaxPool2D.
-
Use of Spatial Excitations.
-
Use of PRelu and Leaky Relus.
Scope of improvement: Yes!
Please feel free to suggest tips for improvement!
-
Use class weights.
-
Use Custom Image Augmentation on Tranining Data.
-
Try other models like PSPNet.
Contribute as much as possible!
Cited by Earth Science Information Partners (ESIP): https://github.com/ESIPFed/Ag-Net