1. Information Technology Research Center, Beijing Academy of Agriculture and Forestry Sciences, Beijing 100097, China
2. National Engineering Research Center for Information Technology in Agriculture, Beijing 100097, China
3. School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, China
Abstract：Automatic early detection of plant diseases is essential for precision crop protection. This paper proposes an early diagnosis and detection method for tomato gray mold based on multi-dimensional spectral series (MDSS) and weighted random forest (WRF) algorithm. The aim was to establish a crop disease detection model by utilizing the overall trend of the spectral curve among multiple observation dimensions of the target leaves to realize the diagnosis before the leaf spot is visible. Generally, the third day after healthy leaves were inoculated with the Botrytis cinerea was treated as the first day that was successfully infected. Therefore, hyperspectral images were recorded from both healthy and infected leaves for 7 days after infection respectively. Then extracted, the region of interest and calculated the average spectrum to form the original spectral samples, whilst (156×7) groups were obtained in total after selection. The group samples were split into multi-dimensional spectral series with 1~7 dimensions per the course of the disease to make up multi-dimensional original spectral series. In order to increase the difference between dimensions, the adjacent original spectral series were subtracted to generate multi-dimensional related spectral series. Afterwards, two symbolic methods, symbolic aggregate approximation (SAX) and symbolic Fourier approximation (SFA), were employed to discretize each spectral series into local discriminant features. Finally, a weighted random forest classification model (MDSS-SAX-SFA-WRF) based on the local discriminant features of multi-dimensional spectral series is established to realize early disease detection. Accordingly, the model based on single-dimensional spectral series (SDSS) is also built as the benchmark to compare with the MDSS-SAX-SFA-WRF model. The experiment results indicate that the MDSS-SAX-SFA-WRF detection model achieves detection accuracies of more than 90% in 56 testing samples containing 2 to 7 spectral series dimensions, and the highest accuracy up to 99% reached in the 5-dimensional sample data, which is 8.2 percentage higher than that of SDSS-SAX-SFA-MRF detection model on the 5th day of infection. Different from the SDSS-SAX-SFA-MRF model detection performance dropped significantly to the lowest 84% in the 5th~7th days of infection due to random interference. While the discrimination accuracy of the MDSS-SAX-SFA-WRF model still retained a high level of more than 98% in the visible stage of infection without excessive decline. Therefore, the classification model based on the overall change trend of the multi-dimensional spectral curve and weighted random forest (MDSS-SAX-SFA-WRF) proposed in this paper can effectively realize the early detection of tomato gray mold with the strong robustness, which provides a new idea for the early differentiation of crop disease.
Key words：Early disease detection; Hyperspectral imaging; Tomato gray mold; Random forest; Multi-dimensional times series
[1] LIU Jie, WANG Fu-xiang, ZENG Juan, et al(刘 杰，王福祥，曾 娟，等). China Plant Protection(中国植保导刊), 2020, 40(7): 5.
[2] QIN Li-feng, ZHANG Xi, ZHANG Xiao-qian(秦立峰，张 熹，张晓茜). Transactions of the Chinese Society for Agricultural Machinery(农业机械学报), 2020, 51(11): 212.
[3] LEI Yu, HAN De-jun, ZENG Qing-dong, et al(雷 雨，韩德俊，曾庆东，等). Transactions of the Chinese Society for Agricultural Machinery(农业机械学报), 2018, 49(5): 226.
[4] YUAN Pei-sen, CAO Yi-fei, MA Qian-li, et al(袁培森, 曹益飞, 马千里，等). Transactions of the Chinese Society for Agricultural Machinery(农业机械学报), 2021, 52(1): 139.
[5] JIN Xiu, QI Hai-jun, LI Shao-wen(金 秀, 齐海军, 李绍稳). Food Science(食品科学), 2018, 39(19): 233.
[6] LIU Si-jia, TIAN You-wen, ZHANG Fang, et al(刘思伽, 田有文, 张 芳, 等). Food Science(食品科学), 2017, 38(8): 277.
[7] Sun Jun, Zhou Xin, Wu Xiaohong, et al. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2019，212：215.
[8] JIA Fang-fang, HONG Quan-chun, SONG Wei-yi(贾方方，洪权春，宋唯一). Chinese Journal of Eco-Agriculture(中国生态农业学报), 2017, 25(6): 805.
[9] Zhang Ning, Yang Guijun, Pan Y C, et al. Remote Sensing, 2020, 12(19): 3188.
[10] Xie C, Yang C, He Y. Computers and Electronics in Agriculture, 2017, 135: 154.
[11] Lucas B, Shifaz A, Pelletier C, et al. Data Mining and Knowledge Discovery, 2019, 33(3): 607.
