Detection of Black Sigatoka Disease on Banana Leaves Using ShuffleNet V2 CNN Architecture in Comparison to SVM and KNN Techniques

Abstract

In this paper, the Shufflenet V2 Convolutional Neural Network Architecture is used to detect Black Sigatoka Disease in banana leaves. This architecture is used to compare its results in terms of accuracy, sensitivity, and specificity with different algorithms that also have been applied to the same scenario. Shufflenet V2 CNN is compared to the Support Vector Machine and K-nearest Neighbor in this case. Image classification has been a helpful tool. Its application detects anomalies and physical manifestations in different cases, such as agriculture and biomedical. Image classification uses different algorithms for its process, and each varies in performance. Thus, this study is made to see the percentage differences in this specific application. The CNN model is trained first by feeding it with data of healthy and Black Sigatoka infected banana leaf images in raw and augmented forms. The trained model is then deployed to a Raspberry Pi device prototype, wherein leaf samples are used as test data. The results of this test garnered 95% accuracy, 96.67% sensitivity, and 93.33% specificity. This ShuffleNet V2 CNN trained model's results are compared to the results of both algorithms, SVM and KNN.