Deep learning assisted continuous wavelet transform-based spectrogram for the detection of chlorophyll content in potato leaves
Introduction
The chlorophyll content of potato crops is an important indicator of photosynthetic capacity and nitrogen status related to its underground tuber (Li et al., 2020a). Traditional chlorophyll content measurement relies on laboratory chemical analysis (e.g., ultraviolet spectrophotometry). Although this method can obtain accurate results, it involves disadvantages of destructive sampling and time-consuming operation (Song et al., 2021a). Visible and near-infrared (Vis-NIR) spectroscopy has proved to be a nondestructive and fast method to estimate the chlorophyll content (Li et al., 2020b, Liu et al., 2018, Zhang et al., 2012) due to the large quantities of features that spectral data contain (Liu et al., 2020a). In spectral analysis, extracting effective features from spectral data is one of the significant issues in chlorophyll content detection. This work may contribute efforts in this topic through monitoring physiological and growth status to indicate field management.
The detection of the chlorophyll content based on Vis-NIR spectroscopy generally depends on features in the bands of blue (400–450 nm), green (centered at 550 nm), red (650–700 nm), and red-edge (700–750 nm) regions, which present the strong absorption and reflection characteristics of chlorophyll (Zhao et al. 2021). The spectral reflectance contains complex information because it is not only determined by the concentration of biochemical components but also crop structural properties and field environment conditions (Ustin et al., 2009, Yu et al., 2014). Thus, finding effective features related to the chlorophyll content is confronted with many interference factors, and it is an obstacle in the accurate detection of the potato leaf chlorophyll content (Yao et al., 2013). Consequently, we focus on identifying effective features sensitive to the leaf chlorophyll content and improving the detection accuracy (Mirzaei et al., 2019).
Many attempts have been made to analyze spectrum features based on signal processing methods, mainly in the spatial domain and time–frequency domain (Berger et al., 2020). Considering spatial domain features, the construction of vegetation indices (VIs) and selection of sensitive wavelengths are used in vegetation monitoring. VI combines a group of fixed wavelengths, and each VI exhibits different suitability in practical applications, which can be classified into the simple ratio vegetation index, normalized difference vegetation index, and soil-adjusted vegetation index (Fu et al., 2021, Xue and Su, 2017). In addition, wavelength selection seeks to identify a subset of spectral variables by the removal of non-informative data when performing quantitative determinations between dissimilar samples (Fu et al., 2021). However, Liu et al. (2019) found that the red edge shifts progressively toward longer wavelengths as the chlorophyll content increases. The feature robustness of the selected wavelengths might be influenced by such a shifting phenomenon, so Li et al. (2014) found that normalized difference red-edge index calculated by reflectance at 720 and 790 nm contributes to the nitrogen estimation results with coefficient of determination (R2) of 0.49 in the V6–V7 corn growth period but 0.79 in the V10–V12 period. Similarly, small changes in the data can lead to different variable selection sets (Yang et al., 2019). Previous research indicated that the features at fixed wavelengths do not effectively present complex spectral differences caused by dynamic growth changes of crops, which might reduce their applicability (Qiao et al., 2020). Thus, identification of complicated features from spectral reflectance is critical to improve the modeling performance of chlorophyll content detection.
By extending the spatial domain into the time–frequency domain, continuous wavelet transform (CWT) has been demonstrated as a potential tool in feature selection and weak information extraction (Zhang et al., 2020). CWT utilizes wavelet functions to decompose the spectral data at different scales into wavelet coefficient curves, which represent the degree of similarity between signals and mother wavelet functions (He et al., 2018). Li et al. (2017) extracted red-edge positions from raw spectrum and wavelet-transformed spectrum to estimate the leaf content of cereal crops. The performance based on CWT achieved better validation results with R2 of 0.86 than raw spectrum with R2 of 0.79. We also applied CWT to detect the chlorophyll content of potato crops (Liu et al., 2020) in primary studies. Liu et al. (2020)) compared the relationship between the wavelet coefficients at different scales and the leaf chlorophyll content; the modeling accuracy at scale 3 performed better (R2 = 0.85) than raw spectrum in the spatial domain (R2 = 0.75). Although previous research proved that CWT can help provide more spectral features related to the chlorophyll content, the wavelengths of wavelet coefficients with higher correlation coefficients shifted among different scales. The changes among scales in CWT might also indicate the features related to the leaf chlorophyll content. However, most studies focused on features at the single scale and did not involve multi-scales when utilizing CWT. Some questions still need to be explored if features among the different scales are useful in the detection of the chlorophyll content. Therefore, it’s necessary to find complicated features considering both scales and wavelengths.
To present comprehensive information of wavelet coefficients at the single scale and changing phenomenon among different scales, we established spectrograms of wavelet coefficients based on CWT, in which we transformed 1D wavelet coefficient curves into 2D spectrograms. Meanwhile, the method of comprehensive feature exploration will be a challenge in spectrogram processing.
Numerous studies have reported that the convolutional neural network (CNN) model is a powerful tool for images or spectra analysis, especially in deeper feature extraction and recognition (Rong et al., 2020, Yang et al., 2019). Weng et al. (2020) applied the CNN model based on Vis-NIR spectroscopy and obtained the best discrimination results with accuracy of 99.00%, but random forest obtained an accuracy of 91.67% in type identification of minced beef adulteration. CNN can extract high-dimensional data features and reduce dimensionality by interleaving convolutional and pooling layers ((Chen et al., 2018;Kattenborn et al., 2021). In view of the spectrogram, Ng et al. (2019) transformed the 1D spectrum into the 2D spectrogram by Hann window to predict several soil properties; the modeling results showed that R2 of the partial least squares model was in the range of 0.87–0.95, and R2 of the CNN model was in the range of 0.90–0.95. These studies indicated that CNN has great potential in spectrogram feature extraction because the deep network helps find more hidden signals (Rong et al., 2020, Yang et al., 2019). Therefore, we proposed a strategy to combine the CNN model with spectrograms based on CWT to explore effective features among wavelengths and multi-scales related to the chlorophyll content.
This study aimed to extract comprehensive features by combining the CNN model and CWT-based spectrogram to improve the accuracy of chlorophyll content detection on the basis of Vis-NIR spectroscopy technology. The main procedures of this study were as follows: (1) compare different preprocessing methods to find appropriate features in the spatial domain; (2) perform CWT to extract robust features in the time–frequency domain; and (3) utilize the CNN model with 2D wavelet coefficient spectrogram to extract comprehensive features and construct the regression model for predicting the chlorophyll content of potato leaves.
Section snippets
Experiments and materials
The experiments were conducted at the national precision agriculture experiment station in Xiao Tangshan town (116°39′ E, 40°17′ N), Beijing, China. The potato cultivar Atlantic was the experimental material used. This work involved 80 plots, and one potato plant per sampling plot was selected. All experimental data were collected at the stem elongation stage (S1), tuber formation stage (S2), tuber bulking stage (S3), and tuber ripening stage (S4); six samples at S1 were discarded. Finally, 314
Chlorophyll content statistics and spectral response of various growth periods
The trend of the mean chlorophyll content with four growth periods is shown in Fig. 4. The potatodataset exhibited a high mean range at the previous growth period (S1 and S2) due to the growth of the aboveground crops. The mean value of the chlorophyll content increased at S3 and S4 because of the formation of underground potato.
The mean spectral curves of four growth periods are shown in Fig. 5. Chlorophyll caused greater reflectance in the green reflection bands (centered at 550 nm) than in
Comparison of preprocessing methods in the spatial domain
As shown in Table 4, the FOD method was employed to transform the raw spectra and improve the model accuracy. The FOD features performed better than the SG features. Derivative is a special form of baseline correction method as it removes constant background signals. It can be used to enhance visual resolution, resolve overlapping peaks, and highlight detailed structures in spectral data. SG smoothing is a widely used pretreatment method that can eliminate noise, such as baseline drift, tilt,
Conclusions
To establish an accurate method based on Vis-NIR spectroscopy for detecting the leaf chlorophyll content of potato crops, this study proposed a feature extraction method combining CWT and the CNN model to extract comprehensive features from 2D wavelet coefficient spectrograms. The accuracy of PLS models constructed by FOD features, wavelet coefficients and spectrogram features by CNN model was compared. The FOD method could enhance the spectral information related to chlorophyll content. CWT can
CRediT authorship contribution statement
Ruomei Zhao: Conceptualization, Methodology, Writing – original draft. Lulu An: Data curation. Weijie Tang: Formal analysis. Dehua Gao: Data acquisition, Formal analysis. Lang Qiao: Data acquisition, Formal analysis. Minzan Li: Methodology, Funding acquisition. Hong Sun: Supervision, Project administration, Funding acquisition. Jinbo Qiao: Investigation, Validation.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This work was supported by the National Natural Science Fund (Grant No. 31971785), the National Key Research and Development Program (Grant No. 2019YFE0125500), University-locality Integrative Development Project of Yantai (2020XDRHXMPT35), and Graduate Student Training Projects of China Agricultural University (JG2019004, YW2020007, JG202026, QYJC202101, JG202102). We would like to acknowledge Ning Liu for his help with field data collection.
References (39)
- et al.
Crop nitrogen monitoring: Recent progress and principal developments in the context of imaging spectroscopy missions
Remote Sens. Environ.
(2020) - et al.
Rapid detection of seven indexes in sheep serum based on Raman spectroscopy combined with DOSC-SPA-PLSR-DS model
Spectrochim. Acta Part A Mol. Biomol. Spectrosc.
(2021) - et al.
Quantitative analysis modeling of infrared spectroscopy based on ensemble convolutional neural networks
Chemometr Intell Lab Syst
(2018) - et al.
An overview of crop nitrogen status assessment using hyperspectral remote sensing: Current status and perspectives
Eur. J. Agron.
(2021) - et al.
Recent advances in convolutional neural networks
Pattern Recogn.
(2018) - et al.
Review on Convolutional Neural Networks (CNN) in vegetation remote sensing
ISPRS J. Photogramm. Remote Sens.
(2021) - et al.
Above-ground biomass estimation and yield prediction in potato by using UAV-based RGB and hyperspectral imaging
ISPRS J. Photogramm. Remote Sens.
(2020) - et al.
WREP: A wavelet-based technique for extracting the red edge position from reflectance spectra for estimating leaf and canopy chlorophyll contents of cereal crops
ISPRS J. Photogramm. Remote Sens.
(2017) - et al.
Improving estimation of summer maize nitrogen status with red edge-based spectral vegetation indices
Field Crops Res.
(2014) - et al.
Lifting wavelet transform for Vis-NIR spectral data optimization to predict wood density
Spectrochim. Acta Part A Mol. Biomol. Spectrosc.
(2020)
Convolutional neural network for simultaneous prediction of several soil properties using visible/near-infrared, mid-infrared, and their combined spectra
Geoderma
Classification of hybrid seeds using near-infrared hyperspectral imaging technology combined with deep learning
Sens. Actuators, B
Using deep learning to predict soil properties from regional spectral data
Geoderma Regional
Peach variety detection using VIS-NIR spectroscopy and deep learning
Comput. Electron. Agric.
Predictors generation by partial least square regression for microwave characterization of dielectric materials
Physica B
Chlorophyll content estimation based on cascade spectral optimizations of interval and wavelength characteristics
Comput. Electron. Agric.
Development of crop chlorophyll detector based on a type of interference filter optical sensor
Comput. Electron. Agric.
Derivative analysis of hyperspectral data
Remote Sens. Environ.
Retrieval of foliar information about plant pigment systems from high resolution spectroscopy
Remote Sens. Environ.
Cited by (21)
Online detection of lycopene content in the two cultivars of tomatoes by multi-point full transmission Vis-NIR spectroscopy
2024, Postharvest Biology and TechnologyClassification of wheat powdery mildew based on hyperspectral: From leaves to canopy
2024, Crop ProtectionHyperspectral imaging detects biological stress of wheat for early diagnosis of crown rot disease
2024, Computers and Electronics in AgricultureVisible and near-infrared spectroscopic determination of sugarcane chlorophyll content using a modified wavelength selection method for multivariate calibration
2024, Spectrochimica Acta - Part A: Molecular and Biomolecular SpectroscopyNondestructive identification of soybean protein in minced chicken meat based on hyperspectral imaging and VGG16-SVM
2024, Journal of Food Composition and AnalysisImproved potato AGB estimates based on UAV RGB and hyperspectral images
2023, Computers and Electronics in Agriculture