Groundnut is an agricultural product that is used for both human consumption and the production of industrial ingredients. However, Groundnut seeds face infected issues by fungi, viruses, pests, physical damage, and high heat. These defects can harm crops in terms of yield, quality, economic value, and health hazards. To address these challenges, computer vision technology is used to classify defects. This study introduced an ensemble deep learning techniques with an averaging strategy of VGG16 and InceptionV3 classification models. The approach utilizes groundnut seed images obtained from trade commodity and agricultural research centers through digital cameras and smartphones. Preprocessing is necessary for images due to variations in brightness, camera light intensity, and size. Preprocessing techniques are selected based on image quality metrics. We then apply watershed segmentation and YOLOv3 detection for region of interest detection and image segmentation. The model development incorporates concatenate of classical and deep-based features. To select suitable feature extraction technique comparison is done between classical feature extracted with (HOG and GLCM) and deep-based feature extraction techniques (InceptionV3 and VGG16) through conducting of three experiments. The classification model was trained with 7800 augmented images with Generative Adversarial Network (GAN) based augmentation, with a split ratio of 10%, 10%, and 80% for the test, validation, and training set. Augmented data and hyper-parameter tuning techniques improved accuracy to 96.25% with an ensemble model. However, misclassifications occurred due to the similar appearance of sample images. Future work should address these challenges and consider oil content and defect grading. Key words:- deep ensemble learning, feature extraction , Generative Adversarial Network, Gray Level Co-occurrence matrix, Histogram of Oriented Gradients, Visual Geometry Group.