Teff, its scientific name Eragrostis Teff is a tiny, round grain that is used to make injera. Possibly Teff is the smallest cereal grain with an average length of ~1mm. East and West Gojjam in Amhara and East and West Shoa in Oromia are known particularly as the Teff producing areas. Teff growing regions in the Amhara region are closer environmental behaviour, thus it is difficult to identify using naked eyes, user can rely upon computer vision techniques to assess complex patterns and accurately identify Teff species. The main objective of this study is designing and developing an automatic Ethiopian Teff identification based on their growing regions using machine-learning techniques. To achieve the objective of the study images of Teff are collected by taking different Teff samples from East and West Gojjam regions of Ethiopia particularly (Adiet, Bichena, Debre Markos, and Dejen). In this study, the four groups of Teff image and each having 772 images are used. That is, form these 3088 images were considered in this study. To capture the image white paper as a background is used to reduce the background noise. Once the dataset is collected, the developed system incorporates different processing phases, such as image resizing, Brightness Preserving Bi Histogram Equalization (BBHE), noise removal, grayscale conversion, contrast enhancement, a combined approach of thresholding and Mask R-CNN segmentation are used as a preprocessing technique in this study. After preprocessing, feature extraction took place. Convolutional neural network (CNN) and Binary Robust Invariant Scalable Key points (BRISK) had been used as feature extraction techniques. To reduce the dimension of the combined features we have applied PCA. In this thesis, multi-class support vector machine (SVM) and CNN are used as a classifier after getting the representing feature vectors. Besides, we have combined CNN with SVM for classifying teff to their predefined growing regions. To enhance the performance of CNN and SVM, Kernel function, optimizers, activation function and learning rate are used as a tuning parameter. For the implementation of the proposed architecture, MATLAB is used as a programming tool. From the experiment, the model achieved the training accuracy of 97.66% and testing accuracy 92.02% achieved. Finally, the confidence interval is used as a validation technique.