Predictive spectral analysis using an end-to-end deep model from hyperspectral images for high-throughput plant phenotyping

• Published in: Computers and electronics in agriculture Vol. 177; p. 105713
• Main Authors: Rehman, Tanzeel U., Ma, Dongdong, Wang, Liangju, Zhang, Libo, Jin, Jian
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.10.2020; Elsevier BV
• Subjects: Artificial neural networks; Convolutional neural networks; Datasets; Deep learning; Hyperspectral imaging; Inception; Infrared cameras; Moisture content; Plant phenotyping; Regression analysis; Spectra; Spectral augmentation; Spectral reflectance; Spectrum analysis; Support vector machines; Deep learning; Spectral augmentation; Inception; Plant phenotyping; Convolutional neural networks; Spectral reflectance; Hyperspectral imaging
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2020.105713

•DeepRWC was developed for predicting the RWC from spectral data of maize plants.•DeepRWC incorporated a modified inception module to extract multi-scale features.•Proposed model outperformed the linear (PLSR) and non-linear (SVR) approaches.•Proposed model can eliminate the need of pre-processing or dimensionality reduction. The spectral reflectance signature of the plants contains rich information about their biophysical, physiological and chemical characteristics. Learning the patterns directly from the plant spectra is critical for predictive plant phenotyping applications. In this study, we developed an end-to-end deep learning model based on 1-D convolutional neural networks, called DeepRWC, to predict the relative water content (RWC) of plants directly from mean spectral reflectance. The proposed model incorporated a modified Inception module to learn multi-scale spectral features at different abstraction levels. To train the proposed network, maize plants grown under well-watered and drought-stressed treatments were imaged using push-broom style, top-view, visible near-infrared (VNIR) hyperspectral camera in the greenhouse environment. Results showed that our proposed model achieved good performance with an R2 of 0.872 for RWC. The performance of the developed model was compared with two standard approaches, partial least squares regression (PLSR) and support vector machine regression (SVR) on two external test datasets. The quantitative analysis showed that the DeepRWC outperformed both linear (PLSR) and non-linear (SVR) approaches by achieving the lowest RMSE and better R2 value on all test datasets included in the study. Our proposed DeepRWC eliminated the need for any preprocessing or dimensionality reduction, as in the case of other standard techniques (PLSR/SVR). These results confirmed the ability of DeepRWC to better predict the RWC of plants using spectral reflectance signature.