Abstract:
Faba Bean is among a crucial crop for countries like Ethiopia due to its high protein content and
high rate of production. It also improves soil fertility with the help of bacteria that forms nodules
on its root. Ethiopia is the world’s second largest producer of Faba Beans next to China however,
the yield is generally low according to researches. Among the constraints which constitute the low
productivity of the crop, disease is the primary one. Even though there are countermeasures, such
as chemical sprays, which can be taken to reduce the damage or eliminate it before harvest, if
possible, an effective disease identification technique is required. Chocolate Spot, Faba Bean Gal,
and Rust are the most common diseases in the Faba Bean growing areas. A total of 8010
representative sample images were used to develop an automatic Faba Bean leaf disease
classification system using a convolutional neural network (CNN) following a design science
research methodology. First, we developed our segmentation algorithm based on Canny edge
detector and removed the image background from our region of interest. Afterward, we have
designed a new CNN network architecture and a modified version of AlexNet architecture to fit
our classification task. In addition, ResNet50V2 and InceptionV3 pre-trained models were chosen
due to their better performance in such tasks. The FBDCNet model is trained with both raw images
and segmented images. When it is trained with raw images an overall testing accuracy of 90%
under the given test set is achieved. Compared to AlexNet, InceptionV3, and ResNet50 the
classification accuracy exceeds by 2%, 6%, and 1% respectively. In the same way training of
FBDCNet model with segmented images give a 98% testing accuracy under the given test set.
Compared to AlexNet, InceptionV3 and ResNet50, the classification accuracy increased by 5%,
2%, and 1% respectively. The experimental results demonstrate that the proposed Canny edge
detection based segmentation technique improves the total testing accuracy of the FBDCNet model
by 8%. Thus, it can be seen that the proposed model effectively identifies Faba Bean leaf diseases
with the proposed segmentation technique. Meanwhile, this study explores a new approach for the
accurate classification of leaf diseases that provides a theoretical foundation for the application of
deep learning in the field of agricultural information.