Early detection of Fusarium wilt in Pepper using multispectral images based on UAV

Abstract

Pepper (Capsicum annuum L.) is an essential seasoning vegetable in Korean food. However, pepper cultivation is constrained by various viruses. Especially, Fusarium wilt is an economic problem threatening pepper production in many countries. The peppers were transplanted on May 2, and the multispectral images were taken on June 28, July 27, and August 26. There were 30 sampling points to measure the vegetation of pepper, but Fusarium wilt infection was confirmed in 15 samples on July 27 and 11 samples on August 25. Therefore, the possibility of Fusarium wilt detection on July 27 and August 25 was confirmed using the multispectral image taken on June 28 and July 27. It was possible to build models for detecting infected peppers using machine learning (KNN; K-Nearest Neighbors, SVM; Support Vector Machine, LR; Logistic Regression) and applying backward elimination to remove the 9 VIs ranked via correlation analysis with the ratio of train and test as 8:2, 7:3, and 6:4. In the case of the disease detection on July 27 using the image of June 28, the KNN model with 8 Vis was selected as the best model with a 7:3 ratio. However, the LR model with NDRE was chosen as the best model for disease detection on July 27 and August 25 using the images of June 28 and July 27 with a 8:2 ratio. The performance of the model which excluded the non-infected samples on August 25 was the best with DVI, TCARI, and RVI as 0.783, 0.733, 0.917, and 0.815 for the calibration and 0.909, 0.833, 1.000, and 0.909 for the validation in order of accuracy, precision, recall, and F1 score. Moreover, there was no error that the infected pepper was confirmed as the non-infected pepper in the convolution matrix. This study aims to develop models for early detection of Pepper Fusarium wilt by calculating vegetation indices based on reflectance values extracted from UAV-based multispectral images and applying them to machine learning classification algorithms. The model developed in this study is expected to contribute to improving the productivity of peppers by preventing the spread of disease through the early detection of pepper wilt.