Crop Classification

With the emergence of artificial intelligence, computer vision, particularly image recognition, has become popular in our society. Image recognition has been applied to identity verification, unlock devices, and payment. 

I fined-tuned a pre-trained convolutional neural network model-VGG16 to classify 5 types of crops, including sugarcane, maize, rice, wheat, and jute. About 1000 photos are trained. 

The model achieves 90% accuracy and can correctly recognize the crop as long as the photo/picture is simply showing only one species and/or focuses on the target crop. When there are more elements (e.g. trees, mountains, chimneys) involved, the model performs poorly. 

More data would definitely be helpful improve model performance but at the same time collecting photos is not easy. I augment data by fliping, rotating, and scaling. However, model performance doesn't improve. This suggests the inclusion of more complex photos/pictures may be useful for model to learn those features.