Scholarly article on topic 'Remote Sensing Technique for Predicting Harvest Time of Tomatoes'

Remote Sensing Technique for Predicting Harvest Time of Tomatoes Academic research paper on "Agriculture, forestry, and fisheries"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Environmental Sciences
OECD Field of science
Keywords
{Tomato / "harvest time" / "visible spectroscopy" / "near-infrared spectroscopy" / "visble-near-infrared spectroscopy" / "partial least- squares regression" / PLSR}

Abstract of research paper on Agriculture, forestry, and fisheries, author of scientific article — Haiqing Yang

Abstract Fast determination of growing stages and harvest time of fruits and vegetables is necessary to implement robotic operation for horticulture automation. This study evaluates the feasibility of using visible-near-infrared (Vis-NIR) spectroscopy to nondestructively determine the harvest time of tomatoes. A mobile, fibre-type, AgroSpec VIS-NIR spectrophotometer (Tec5, Germany) with a spectral range of 350-2200nm, was used for spectral acquisition of tomatoes in reflection mode. A new index was used to measure the growing stages of tomatoes. Tomato plants were provided by Silsoe Horticultural Center, Bedfordshire, United Kingdom. Spectra were divided into a calibration set (70%) and an independent validation set (30%). Calibration set were subjected to a partial least squares regression (PLSR) with leave-one-out cross validation to establish calibration models respectively based on different spectral ranges, e.g., VIS(400-760nm), NIR(760-2100nm) and VIS-NIR(400-2100nm). Prediction performance of these models on the independent validation set indicates that PLSR models based on entire spectral range (VIS-NIR) outperform those based on partial spectral ranges (VIS or NIR). Coupled with appropriate spectral transformation, the PLSR models can achieve excellent prediction performance of harvest time of tomatoes with coefficient of determination (R 2) of 0.89 and RPD of 3.00. It is concluded that VIS-NIR spectroscopy combined with optimized PLSR models for GS prediction can be successfully adopted as a remote sensing technique for predicting harvest time of tomatoes, which allows for implementing autonomous fruit-picking robots.

Academic research paper on topic "Remote Sensing Technique for Predicting Harvest Time of Tomatoes"

Available online at www.sciencedirect.com

SciVerse ScienceDirect PrOCSd ¡0

Environmental Sciences

Procedia Environmental Sciences 10 (2011) 666 - 671

2011 3rd International Conference on Environmental Science and Information Application Technology (ESIAT 20 11)

Remote Sensing Technique for Predicting Harvest Time of

Tomatoes

Haiqing Yang

College of Information Engineering, Zhejiang University of Technology, Hangzhou 310032, P.R.China

Abstract

Fast determination of growing stages and harvest time of fruits and vegetables is necessary to implement robotic operation for horticulture automation. This study evaluates the feasibility of using visible-near-infrared (Vis-NIR) spectroscopy to nondestructively determine the harvest time of tomatoes. A mobile, fibre-type, AgroSpec VIS-NIR spectrophotometer (Tec5, Germany) with a spectral range of 350-2200nm, was used for spectral acquisition of tomatoes in reflection mode. A new index was used to measure the growing stages of tomatoes. Tomato plants were provided by Silsoe Horticultural Center, Bedfordshire, United Kingdom. Spectra were divided into a calibration set (70%) and an independent validation set (30%). Calibration set were subjected to a partial least squares regression (PLSR) with leave-one-out cross validation to establish calibration models respectively based on different spectral ranges, e.g., VIS(400-760nm), NIR(760-2100nm) and VIS-NIR(400-2100nm). Prediction performance of these models on the independent validation set indicates that PLSR models based on entire spectral range (VIS-NIR) outperform those based on partial spectral ranges (VIS or NIR). Coupled with appropriate spectral transformation, the PLSR models can achieve excellent prediction performance of harvest time of tomatoes with coefficient of determination (R2) of 0.89 and RPD of 3.00. It is concluded that VIS-NIR spectroscopy combined with optimized PLSR models for GS prediction can be successfully adopted as a remote sensing technique for predicting harvest time of tomatoes, which allows for implementing autonomous fruit-picking robots.

© 2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of Conference ES IAT2011 Organizat ion Committee.

Keywords: Tomato; harvest time; visible spectroscopy; near-infrared spectroscopy; visble-near-infrared spectroscopy; partial least-squares regression; PLSR

1. Introduction

In the previous study[1], a new growing stage (GS) index was proposed with the purpose of predicting harvest time of tomatoes of three cultivars. However, there is no information about the optimation of

* Corresponding author. Tel.: +86-571-56330001; fax: +86-571-56330001. E-mail:yanghq@zjut.edu.cn.

1878-0296 © 2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of Conference ESIAT2011 Organization Committee. doi: 10. 1016/j .proenv .2011.09. 107

spectral range on which calibration models were developed. The issue of whether the visible range (VIS:400-760nm) or near infrared range (NIR:760+ nm) is adequate for building accurate calibration models should be examined.

The objective of this study is to evaluate the influence of different spectral ranges on the prediction accuracy of calibration models for predicting harvest time of tomatoes.

2. Materials and methods

2.1. Tomato samples

Three cultivars of tomato plants were cultivated at the Silsoe Horticultural Centre, Bedfordshire, United Kingdom, during summer growing season in 2010. A total of 75 fruits were selected for the study. Instrumental measurement started on the 24th July and continued with an interval of 2-3 days until the targeted tomatoes were fully ripe (more than 90% red) and picked based on the USDA Tomato Ripeness Color Chart. The statistics of tomatoes were listed in Table 1.

Table 1 Sample statistics of tomatoes

Data set Fruit Spectral Picking Date GS index

Number Number First Last Mean Range SD

Calibration 60 654 05 Aug 06Sep 0.452 0-1 0.310

Validation 15 165 07Aug 06Sep 0.447 0-1 0.303

2.2. Spectral Acquisition

The reflectance spectra of tomatoes were measured with a mobile fibre-type AgroSpec VIS-NIR spectrophotometer (Tec5, Germany) with spectral region of 350-2200nm. A USB cable was used for the data transmission between the spectrophotometer and a portable computer. A 100% white reference was used before scanning. Measurement was made at three separate positions on the equator of a fruit. A total of ten scans were measured at each position and the spectra from the three positions were averaged as one sample.

2.3. Tomato growing stage (GS) index

In order to build a uniformed calibration model, a new growing stage (GS) index was defined as the ratio of the current growing age in days to the on-vine duration before harvest of tomatoes in days. The details were introduced in the published report[1].

2.4. Spectral transformation

The spectra were divided into a calibration set (70%) and an independent validation set (30%). The calibration spectra were subjected to a partial least squares regression (PLSR) with leave-one-out cross validation. The PLSR relates the variations in one response variable (GS) to the variations of several predictors (wavelengths). The optimal number of latent variables (LVs) was determined by minimizing the predicted residual error sum of squares (PRESS).

To investigate the influence of different spectral ranges on model performance, calibration models were established respectively based on the entire spectral range of 400-2100nm (VIS-NIR) and partial ranges of 400-760nm (VIS) and 760-2100nm (NIR). Meanwhile, spectral transformation was conducted using several algorithms including baseline offset correction (BOC), 1st and 2nd order de-trendings, and 1st

and 2nd derivatives. The performance of models for transformed spectra was compared to those for original spectra.

The PLSR models were evaluated using coefficient of determination (R2) in calibration and cross validation, root-mean-square error of calibration (RMSEC) and cross validation (RMSECV). The performance of the PLSR models on the independent validation set was assessed using the coefficient of determination (R2), the root-mean-square error of prediction (RMSEP) and the residual prediction deviation (RPD. We adopted the criteria of classifying RPD values [2] as follows: an RPD value below 1.5 indicates that the calibration is not usable; an RPD value between 1.5 and 2.0 indicates a possibility to distinguish between high and low values; an RPD value between 2.0 and 2.5 makes approximate quantitative predictions possible. For RPD value between 2.5 and 3.0 and above 3.0, the prediction is classified as good and excellent, respectively. Generally, a good model should have high values of R2 and RPD, and low values of RMSEC, RMSECV and RMSEP. The spectral transformation and PLSR calibration were conducted using the Unscrambler® (CAMO Software AS, Oslo, Norway).

3. Results and discussion

3.1. Spectral characteristics

The characteristic reflectance spectra of tomato at different growing stages are illustrated in Fig.1. The

significant wavelengths with spectroscopic explanations are listed in Table 1.

0.9 - -24Jul — 27Jul .....29Jul

0.8 - r' \ -----31 Jul ......03Aug .......05Aug

0.7 p1 \ *1 \ -07 Au g ------10Aug -12Aug

i ' \ r i \ — 14Aug .....17Aug -----20Aug

0.6 \ \ ------2 5 Au g .......28Aug --31 Aug

» \ -— 03Sep -06Sep

_i_i_i_i_i_i_i_i_i_

200 400 600 800 1000 1200 1400 1600 1800 2000 2200

Wavelength(nm)

Fig.1. Spectral characteristics of a representative of tomatoes at different growing stages[1]

Table 1 Significant wavelengths and spectroscopic explanations

Wavelength (nm) Spectroscopic explanation Reference

491 lycopene and P-carotene [3]

672 absorption of chlorophyll a [4, 5]

763 the 3rd overtone of O-H stretching [3]

981 the 3rd overtone of H2O stretching and bending absorbance [3]

1204 sugar absorption [7]

1456 the 2nd overtones of H2O stretching and bending, respectively [6]

1928 the 1st overtones of H2O stretching and bending, respectively [6]

3.2. PLSR models on VIS range (400-760nm)

Table 2 reports the result of PLSR models for original and transformed spectra based on VIS range (400-760nm). In general, these models are only useful as approximate quantitative predictions according to the RPD classification. The best prediction was obtained for the original spectra with R2 and RPD values of 0.84 and 2.51, respectively.

Table 2 PLSR models based on VIS range (400-760nm)

Spectral transformation LVs Calibration Cross-validation Independent validation

R2 RMSEC R2 RMSECV R2 RMSEP Bias RPD

None 7 0.84 0.123 0.83 0.126 0.84 0.121 -0.004 2.51

BOC 7 0.83 0.126 0.82 0.130 0.82 0.129 -0.001 2.35

1st De-trending 5 0.81 0.134 0.80 0.139 0.81 0.133 -0.002 2.27

2nd De-trending 7 0.84 0.125 0.83 0.129 0.83 0.124 -0.003 2.44

1stderivative 8 0.85 0.119 0.84 0.125 0.83 0.125 0.002 2.43

3.3. PLSR models on NIR range (760-2100nm)

Table 3 reports the result of PLSR models for original and transformed spectra based on NIR range (760-2100nm). In general, these models are not suitable for quantitative predictions, although the performance of PLSR model can be improved by appropriate spectral transformation.

Table 3 PLSR models based on VIS range (760-2100nm)

Spectral transformation LVs Calibration Cross-validation Independent validation

R2 RMSEC R2 RMSECV R2 RMSEP Bias RPD

None 9 0.81 0.133 0.80 0.139 0.81 0.131 0.010 2.29

BOC 9 0.83 0.127 0.81 0.134 0.82 0.130 0.010 2.31

1st De-trending 7 0.82 0.132 0.81 0.136 0.82 0.129 0.012 2.32

2nd De-trending 7 0.81 0.135 0.80 0.138 0.81 0.131 0.008 2.29

1st derivative 5 0.83 0.129 0.79 0.141 0.77 0.145 0.009 2.07

3.3. PLSR models on VIS-NIR range (400-2100nm)

Table 4 reports the result of the PLSR models for original and transformed spectra based on entire spectral range (400-2100nm). By comparison, these models outperform those based on partial spectral ranges (VIS or NIR). It suggests that neither VIS nor NIR ranges include all spectroscopic information for predicting harvest time of tomatoes. This can be confirmed by the distribution of significant wavelengths across the range of 400-2100nm (Fig.1, Table 1). The performance of these models can be regarded as good or excellent, although the accuracy of these PLSR models is a function of spectral pretreatment. For

instance, the best prediction accuracy was obtained for the spectra transformed by 2nd de-trending with R2 and RPD values of 0.89 and 3.00, respectively.

Table 4 PLSR models based on VIS-NIR range (400-2500nm)

Spectral transformation LVs Calibration Cross-validation Independent validation

R2 RMSEC R2 RMSECV R2 RMSEP Bias RPD

None 7 0.88 0.108 0.87 0.111 0.86 0.112 0.012 2.73

BOC 8 0.88 0.107 0.88 0.109 0.87 0.109 0.012 2.80

1st De-trending 7 0.88 0.109 0.87 0.112 0.87 0.108 0.008 2.81

2nd De-trending 9 0.90 0.100 0.88 0.105 0.89 0.101 0.007 3.00

1stderivative 8 0.90 0.097 0.89 0.102 0.89 0.102 0.009 2.97

3.4. Influential wavelengths contributing to PLSR calibration models

In order to identify those wavelengths across the range of 400-2100 nm, where spectral reflectance had intensively influenced on the model development, PLSR coefficient curve obtained from GS calibration with optimized 7 LVs were analyzed. Figure 2 shows the influential wavelengths on the calibration models for GS index. For examples, a negative band at 516 nm, possibly associated with either lycopene or beta-carotene [8,9], exhibits strong influence on GS prediction. In the NIR range, more influential bands can be found at 934 nm (3rd overtone of H2O vibration), at 1120 nm (sugar-related absorption band), at around 1168 and 1335 nm (2nd overtone of O-H stretch of carboxylic acid dimmers and/or 2nd overtone of C-H stretch of CH3 groups) [6], and at around 1719 nm (3rd overtone of C-O stretch of amino acid ionized carbonyls and/or 3rd overtone of P=O hydrogen bonded) [6]. All these wavelengths present stronger influence on GS prediction than other spectral bands.

Wavelength(nm)

Fig.2 Important wavelengths marked in the PLSR coefficient curve

4. Conclusions

The visible-near-infrared (VIS-NIR) spectroscopy was used for determining harvest time of tomatoes of three cultivars. Calibration models were established by a partial least squares regression (PLSR) for original and transformed spectra. Model performance was compared among different spectral ranges, e.g. VIS (400-760nm), NIR(760-2100nm) and VIS-NIR(400-2100nm). Validation result shows that PLSR models on VIS-NIR range outperform those on VIS or NIR ranges. It suggests that neither VIS nor NIR ranges include all spectroscopic information for predicting harvest time of tomatoes. It is concluded that VIS-NIR spectroscopy combined with optimized PLSR models for GS prediction can be successfully adopted as a remote sensing technique for predicting harvest time of tomatoes, which allows for implementing autonomous fruit-picking robots.

Acknowledgements

The author gratefully thanks the Silsoe Horticultural Center, Bedfordshire, United Kingdom for the provision of tomato plants. Specially thanks to Dr. Abdul M. Mouazen and Mr. Boyan Kuang for providing spectrophotometer and in-field experimental assistance. This study was financially supported by the Natural Science Foundation of Zhejiang Province, P.R.China (No. Y1090885) and by the State Scholarship Fund of China (Grant No.[2009]3004).

References

[1] Yang H, Kuang B, Mouazen AM. In situ determination of growing stages and harvest time of tomato (Lycopersicon esculentum) fruits using fiber-optic visible-near-infrared (Vis-NIR) spectroscopy. Applied Spectroscopy 2011;65:49-56.

[2] Mouazen AM, Baerdemaeker JD, Ramon H. Effect of wavelength range on the measurement accuracy of some selected soil constituents using visual-near infrared spectroscopy. Journal of Near Infrared Spectroscopy 2006;14:189-99.

[3] Clément A, Dorais M, Vernon M. Multivariate Approach to the Measurement of Tomato Maturity and Gustatory Attributes and Their Rapid Assessment by Vis-NIR Spectroscopy. Journal of Agricultural & Food Chemistry 2008;56:1538-44.

[4] Tijskens LMM, Zerbini PE, Schouten RE, Vanoli M, Jacob S, Grassi M, Cubeddu R, Spinelli L, Torricelli A. Assessing harvest maturity in nectarines. PostharvestBiology and Technology 2007;45:204-13.

[5] Ziosi V, Noferini M, Fiori G, Tadiello A, Trainotti L, Casadoro G, Costa G. A new index based on vis spectroscopy to characterize the progression of ripening in peach fruit. Postharvest Biology and Technology 2008;49:319-29.

[6] Murry I, Williams PC. Chemical Principals of Near-infrared technology. In: Williams PC, Norris KH (eds.) AACC, St. Paul, MN, 1987:17-34.

[7] Pérez-Marín D, Sánchez MT, Paz P, Soriano MA, Guerrero JE, Garrido-Varo A. Non-destructive determination of quality parameters in nectarines during on-tree ripening and postharvest storage. Postharvest Biology and Technology 2009;52:180-8.

[8] Clément A, Dorais M, Vernon M. Nondestructive Measurement of Fresh Tomato Lycopene Content and Other Physicochemical Characteristics Using Visible-NIR Spectroscopy. Journal of Agricultural & Food Chemistry 2008;56: 9813-8.

[9] Lai A, Santangelo E, Soressi GP, Fantoni R. Analysis of the main secondary metabolites produced in tomato (Lycopersicon esculentum, Mill.) epicarp tissue during fruit ripening using fluorescence techniques. Postharvest Biology and Technology 2007; 43:335-42.