Integrating spectral and image data to detect Fusarium head blight of wheat

• Published in: Computers and electronics in agriculture 2020-08, Vol.175, p.105588, Article 105588
• Main Authors: Zhang, Dong-Yan, Chen, Gao, Yin, Xun, Hu, Rong-Jie, Gu, Chun-Yan, Pan, Zheng-Gao, Zhou, Xin-Gen, Chen, Yu
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Blight; Color; Convolutional neural network; Datasets; Decision trees; Feature extraction; Fusarium; Fusarium head blight; Fusion of spectral and image; Hyperspectral imaging; Image acquisition; Prediction model; Principal components analysis; Random forest; Spectra; Wavelet transforms; Wheat
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2020.105588

•Our decision tree method selected optimal spectral and image features of wheat Fusarium head blight.•Spectral and image data were integrated into fusion feature.•A deep learning model performed best for detecting wheat Fusarium head blight. Fusarium head blight (FHB), caused by the fungus Gibberella zeae, infects spikelets on wheat heads and can cause significant yield and quality losses in wheat. Application of hyperspectral imaging on the detection of FHB was evaluated in the current study. Hyperspectral images were acquired from a total of 1,680 Fusarium-infected wheat head samples over a wavelength range of 400–1000 nm. The principal component analysis was used to reduce dimension of the hyperspectral image. The central wavelengths at 660, 560 and 480 nm were combined into the RGB image and then transferred to YDbDr space. The texture features of the first six principal components were extracted based on gray level co-occurrence matrix and dual-tree complex wavelet transform and the color features extracted in the color space of RGB and YDbDr. Gradient boosting decision tree and sequential backward elimination were applied to select the optimal features, and 50 spectral features and 40 image features were screened. The random forest model was built based on spectral, image, and fusion features of both spectral and image features of wheat heads to determine the optimal features dataset. Then, the deep convolutional neural network (DCNN) was established based on the optimal features dataset. This process resulted in the development of the DCNN model that predicted disease severity most accurately (R2 = 0.97 and RMSE = 3.78). The DCNN model developed from this study can be used as a new tool to detect and predict the FHB disease in wheat.