Faba bean is one of the most important agricultural products traded internationally and used for domestic consumption in Ethiopian feeding culture. Ethiopia is the second faba bean producing country in the world next to the Republic of china. The variety attributes of faba bean are the parameters that producers are most concerned about, and different faba bean varieties have different quality attributes, which also make faba bean variety identification an important issue in this research. Although traditional machine learning approach have been used for different variety identification using global features, it is sensitive to variability in terms of image transformation or object appearance and it also hampers the performance of the model. This calls for the development of an automatic identification of faba bean variety using different local descriptors and deep learning approach. The proposed model is implemented using Keras (using TensorFlow as a backend) in Python and a sample dataset collected from Adet Agricultural Center. The proposed model has four components: preprocessing, segmentation, feature extraction and identification. In image preprocessing, we normalize the image to improve the image by suppressing unnecessary distortions. Segmentation is used for region of interest extraction and we have used an Otsu and watershed segmentation algorithm to extract the ROI (beans) of the image. In feature extraction, we have followed two approaches to identify faba bean variety. In the first approach, we have investigated the impact of combining different local descriptor features (HOG, SURF, LBP, and GLCM) on the performance of faba bean variety identification and PCA dimensionality reduction methods to train RF, SVM and MLP classifier. In the second approach, we have used convolutional neural networks where the model trained with adaptively learnt features and their levels to identify images. We also investigate the impact of aggregating various local descriptors with CNN models which achieved a state of art results when the dataset becomes small. From the overall experimentation, the proposed end to end CNN model achieved 98.67% training accuracy and testing 92.46 % accuracy by increasing the number of dataset.