AgriculturalNet-AgNet

State of the art model for classifying 256 crop categories built in Keras.

Basic Architecture Used:

U-Net

Dataset:

https://github.com/ZihengSun/Ag-Net-Dataset

Color Map:

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%.

True Image

Predicted Image

Changes to U-Net block:

1. Use of skip connections.

2. Use of GlobalMaxPool2D instead of MaxPool2D.

3. Use of Spatial Excitations.

4. Use of PRelu and Leaky Relus.

Scope of improvement: Yes!

Please feel free to suggest tips for improvement!

Suggested Tips:

1. Use class weights.

2. Use Custom Image Augmentation on Tranining Data.

3. Try other models like PSPNet.

Contribute as much as possible!

Cited by Earth Science Information Partners (ESIP): https://github.com/ESIPFed/Ag-Net