Browsing by Author "Elinisa, Christina"

Now showing 1 - 1 of 1

Image segmentation deep learning model used to Identify black Saigatoka And Fusarium wilt in banana early
(NM-AIST, 2024-05) Elinisa, Christina
Bananas are among the most widely produced perennial fruit crops. Farmers largely produce bananas because they are important staple food and cash crops. However, bananas are highly affected by Fusarium Wilt and Black Sigatoka diseases. These diseases cause yield losses ranging from 30% to 100% of all the banana produce. Farmers face challenges in detecting and mitigating the effects of these two banana diseases because of a lack of knowledge of the diseases and the use of traditional eye observation method in detection. This study is inspired by the success of deep learning and computer vision in detecting a wide range of plant diseases. The study proposed the use of deep learning to automate the early detection of Fusarium Wilt and Black Sigatoka banana diseases. Mask R-CNN and U-Net image segmentation deep learning models were assessed for instance and semantic image segmentation. A dataset comprising 27 360 images of banana leaves and stalks that are healthy, Fusarium Wilt infected, and Black Sigatoka infected collected from the farm was used to train the models. An addition of 407 images of other things apart from the banana plant were downloaded from the internet and used to train the CNN model. From the experiments, the Mask R-CNN model achieved a mean Average Precision of 0.045 29 in segmenting the two banana diseases. The U-Net model achieved an Intersection over Union (IoU) of 93.23% and a Dice Coefficient of 96.45%. Similar results were obtained by Loyani et al. (2021) when they segmented a tomato plant paste called tuta absoluta using a U-Net model. Their model achieved a Dice Coefficient of 82.86% and an Intersection over Union of 78.60%. Additionally, the Fusarium Wilt and Black Sigatoka infected banana leaves and stalks were segmented using the Mask R-CNN and U-Net models. The CNN model yielded an accuracy of 91.71% in classifying the two banana diseases. Similar results were obtained by Sanga et al. (2020) when they deployed an Inceptionv3 model, which achieved an accuracy of 95.41%. The CNN model was deployed in a mobile application to be used by farmers to detect the two banana diseases early. The mobile application could detect banana diseases early and provide research-based mitigation recommendations that smallholder farmers and other agricultural stakeholders can use to avoid yield losses and financial losses.