Understanding the learning mechanism of convolutional neural networks in spectral analysis

Zhang, Xiaolei; Xu, Jinfan; Yang, Jie; Chen, Li; Zhou, Haibo; Liu, Xiangjiang; Li, Haifeng; Lin, Zhixian; Ying, Yibin

doi:10.1016/j.aca.2020.03.055

Cited by 94 publications

(45 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…1). Machine learning models tend to reach a plateau or show marginal improvement with an increasing amount of data, as the model has limited complexity to deal with an increasing amount of data (Zhu et al, 2016).…”

Section: Introductionmentioning

confidence: 99%

The influence of training sample size on the accuracy of deep learning models for the prediction of soil properties with near-infrared spectroscopy data

Minasny

Mendes³

et al. 2020

SOIL

109

View full text Add to dashboard Cite

Abstract. The number of samples used in the calibration data set affects the quality of the generated predictive models using visible, near and shortwave infrared (VIS–NIR–SWIR) spectroscopy for soil attributes. Recently, the convolutional neural network (CNN) has been regarded as a highly accurate model for predicting soil properties on a large database. However, it has not yet been ascertained how large the sample size should be for CNN model to be effective. This paper investigates the effect of the training sample size on the accuracy of deep learning and machine learning models. It aims at providing an estimate of how many calibration samples are needed to improve the model performance of soil properties predictions with CNN as compared to conventional machine learning models. In addition, this paper also looks at a way to interpret the CNN models, which are commonly labelled as a black box. It is hypothesised that the performance of machine learning models will increase with an increasing number of training samples, but it will plateau when it reaches a certain number, while the performance of CNN will keep improving. The performances of two machine learning models (partial least squares regression – PLSR; Cubist) are compared against the CNN model. A VIS–NIR–SWIR spectra library from Brazil, containing 4251 unique sites with averages of two to three samples per depth (a total of 12 044 samples), was divided into calibration (3188 sites) and validation (1063 sites) sets. A subset of the calibration data set was then created to represent a smaller calibration data set ranging from 125, 300, 500, 1000, 1500, 2000, 2500 and 2700 unique sites, which is equivalent to a sample size of approximately 350, 840, 1400, 2800, 4200, 5600, 7000 and 7650. All three models (PLSR, Cubist and CNN) were generated for each sample size of the unique sites for the prediction of five different soil properties, i.e. cation exchange capacity, organic carbon, sand, silt and clay content. These calibration subset sampling processes and modelling were repeated 10 times to provide a better representation of the model performances. Learning curves showed that the accuracy increased with an increasing number of training samples. At a lower number of samples (< 1000), PLSR and Cubist performed better than CNN. The performance of CNN outweighed the PLSR and Cubist model at a sample size of 1500 and 1800, respectively. It can be recommended that deep learning is most efficient for spectra modelling for sample sizes above 2000. The accuracy of the PLSR and Cubist model seems to reach a plateau above sample sizes of 4200 and 5000, respectively, while the accuracy of CNN has not plateaued. A sensitivity analysis of the CNN model demonstrated its ability to determine important wavelengths region that affected the predictions of various soil attributes.

show abstract

Section: Introductionmentioning

confidence: 99%

The influence of training sample size on the accuracy of deep learning models for the prediction of soil properties with near-infrared spectroscopy data

Minasny

Mendes³

et al. 2020

SOIL

109

View full text Add to dashboard Cite

show abstract

“…The CNN model has been well applied in image recognition (Zhao, Hou, Lv, Zhu, & Ding, 2020), hyperspectral image identification (Qiu et al, 2018), spectral analysis (Yang et al, 2019), and so forth. Comparing with conventional spectral analysis methods, effective features from complex multi-dimensional data can be automatically selected and extracted (X. L. Zhang et al, 2020). There, however, has few studies considering the increase of the one-dimensional CNN (1D-CNN) model width to fuse deep spectral features for a better identification of pesticide residues based on Vis/NIR spectroscopy.…”

mentioning

confidence: 99%

Nondestructive identification of pesticide residues on the Hami melon surface using deep feature fusion by Vis/NIR spectroscopy and 1D‐CNN

Chen

et al. 2020

J Food Process Engineering

View full text Add to dashboard Cite

Nondestructive identification of pesticide residues remains a challenge in terms of fruit safety assessment. In this study, a novel method based on visible/near‐infrared (Vis/NIR) spectroscopy (348.45–1,141.34 nm) combined with deep feature fusion was proposed, achieving nondestructive identification of pesticide residues on the Hami melon surface. The spectra of Hami melons with clear water and three kinds of pesticide residues (chlorothalonil, imidacloprid, and pyraclostrobin) were collected in the diffuse reflectance mode. The one‐dimensional convolutional neural network (1D‐CNN), with increased width and depth through parallel convolution modules and concatenate layers, was presented to capture multiple deep features from Vis/NIR spectra and fuse them. This model had a better performance for four‐class identification as the accuracy of 95.83%, and outperformed other CNN models and conventional approaches (partial least squares discriminant analysis and support vector machine). Moreover, the proposed 1D‐CNN model could accurately differentiate whether there were pesticide residues with the identification accuracy as 99.17%. However, the prediction of imidacloprid and pyraclostrobin residues was not accurate due to the similar spectral features. The overall studies indicated that the 1D‐CNN model with deep feature fusion looked promising for nondestructive identification of pesticide residues on the Hami melon surface based on Vis/NIR spectroscopy.Practical applicationsVisible and near‐infrared (Vis/NIR) spectroscopy, as a nondestructive technique, looks promising for evaluation of fruit quality and safety. One‐dimensional convolutional neural network, with deep feature fusion structure to capture multi‐scale spectral information, has a better identification of pesticide residues on the Hami melon surface. Vis/NIR spectroscopy with deep feature fusion can be applied in research and development of a nondestructive detector for pesticide residues on the thick‐skinned fruit surface in the future.

show abstract

“…CNNs create hierarchical representations of data showing how specific irrelevant spectral features are discarded and specific areas of the spectra needed to predict DBP concentrations are magnified. 29 The initial filter layer identifies large and smooth and broad features in the spectra. After pooling, feature maps become more coarse and more distinct patterns between filters can be discerned, highlighting specific areas of the spectra.…”

Section: Convolutional Networkmentioning

confidence: 99%

“…28 Furthermore, CNNs typically employ pooling layers where outputs in specific locations are merged with nearby outputs, creating invariance to small distortions in the input and reducing the dimensionality of the representation. 27,28 CNNs have been successfully applied in chemometric applications such as interpreting Raman and mid-infrared spectra for identifying Escherichia coli and meats, 29 pharmaceuticals in tablets with near infrared spectra, 30 categorizing wines using infrared spectra, 31 and classification of manganese valence. 32 However, there has been no use of CNNs for interpreting 2D fluorescence spectra, and previous implementations have focused on 1D infrared or Raman spectra.…”

Section: Introductionmentioning

confidence: 99%

Application of Convolutional Neural Networks for Prediction of 1 Disinfection By-Products

Peleato

2021

Preprint

View full text Add to dashboard Cite

Fluorescence spectroscopy can provide high-level chemical characterization and quantification that is suitable for use in online process monitoring and control. However, the high-dimensionality of excitation-emission matrices and superposition of underlying signals is a major challenge to implementation. Herein the use of Convolutional Neural Networks (CNNs) is investigated to interpret fluorescence spectra and predict the formation of disinfection by-products during drinking water treatment. Using deep CNNs, mean absolute prediction error on a test set of data for total trihalomethanes, total haloacetic acids, and the major individual species were all < 6 µg/L and represent a significant difference improved by 39% - 62% compared to dense neural networks. Heat maps that identify spectral areas of importance for prediction showed unique humic-like and protein-like regions for individual disinfection by-product species that can be used to validate models and provide insight into precursor characteristics. The use of fluorescence spectroscopy coupled with deep CNNs shows promise to be used for rapid estimation of DBP formation potentials without the need for extensive data pre-processing or dimensionality reduction. Knowledge of DBP formation potentials in near real-time can enable tighter treatment controls and management efforts to minimize the exposure of the public to DBPs.

show abstract

Understanding the learning mechanism of convolutional neural networks in spectral analysis

Cited by 94 publications

References 23 publications

The influence of training sample size on the accuracy of deep learning models for the prediction of soil properties with near-infrared spectroscopy data

The influence of training sample size on the accuracy of deep learning models for the prediction of soil properties with near-infrared spectroscopy data

Nondestructive identification of pesticide residues on the Hami melon surface using deep feature fusion by Vis/NIR spectroscopy and 1D‐CNN

Application of Convolutional Neural Networks for Prediction of 1 Disinfection By-Products

Contact Info

Product

Resources

About