Original papers
Self-organizing map estimator for the crop water stress index

https://doi.org/10.1016/j.compag.2021.106232Get rights and content

Highlights

  • SOM based model is developed for predicting the CWSI using environmental variables.

  • Predicted CWSI correlated well with the observed CWSI for common field moisture conditions.

  • SOM model performance is impacted by significant proportion of zero CWSI values in the dataset.

Abstract

Crop water stress index (CWSI) is a reliable, economic and non-destructive method of monitoring the onset of water stress for irrigation scheduling purposes. Its application, however, is limited due to the need of obtaining the baseline canopy temperatures. This study developed a self-organizing map (SOM) based model to predict the CWSI using microclimatic variables, namely air temperature, canopy temperature and relative humidity. The canopy temperature measurements were made from Indian mustard crop grown in a humid sub-tropical agro-climate during the 2017 and 2018 cropping seasons. Eight levels of irrigation treatments (I1 – I8) based on maximum allowable depletion of available soil water were considered in the study. The CWSI for treatments I2 – I7 was computed using the empirical approach based on the experimentally measured baseline canopy temperatures from treatments I1 and I8. The number of data points used was 1260 and 1350 for model training and testing, respectively. The developed SOM model was evaluated using the error indices Nash-Sutcliffe efficiency (NSE), bias error (BE), absolute error (AE), and coefficient of determination (R2). The SOM predicted CWSI presented a good agreement with the baseline computed CWSI values during model training (R2 = 0.98, NSE = 0.97, AE = 0.018, BE = 0.0004) and testing (R2 = 0.98, NSE = 0.98, AE = 0.018, BE = 0.002). Treatment specific analysis was conducted to evaluate the performance of SOM predicted CWSI for different irrigation levels. Results indicated that the presence of zero CWSI values in a significant proportion in the dataset impacted the model prediction performance at low CWSI (<0.1) values, with an R2 of 0.71 during testing. Nonetheless, the model performed exceptionally well in predicting CWSI values between 0.1 and 0.6 (R2 = 0.93–0.98, NSE = 0.92–0.98, AE = 0.013–0.015, BE = −0.002–0.004), which is the commonly observed CWSI range for irrigation scheduling in field crops. For better understanding, the developed SOM model was also analysed through the component planes, U-matrix, clusters and high-low bar planes in the cluster features.

Introduction

Recent advances in irrigation scheduling have stressed upon the need of remote sensing methods for efficient, precise and sustainable management of agricultural water (Ihuoma and Madramootoo, 2017, Gerhards et al., 2019, Weiss et al., 2020). The use of infrared thermometers (IRT) for measuring the leaf temperatures is a widely acknowledged approach for monitoring plant water status and detecting the onset of water stresses (Kumar et al., 2020b). Crop water stress index (CWSI), initially proposed by Jackson et al., 1981, Idso et al., 1981, is one such method which employs the IRT based canopy temperature (Tcanopy) measurements and other environmental variables including air temperature (Tair), solar radiation, wind speed and relative humidity (RH) (Payero and Irmak, 2006). CWSI has gained a wide reputation and significant advances have been reported in the last two–three decades owing to its direct, reliable and non-destructive nature (Osroosh et al., 2016, Virnodkar et al., 2020).

The principle utilized in the development of CWSI is the inverse relation of Tcanopy with leaf transpiration and stomatal closure (Ehrler, 1973). Due to water stress in plants, the stomatal closure takes place causing a reduction in transpirational cooling, thereby increasing Tcanopy, which acts as an indicator for stress (Fuchs, 1990, Kumar et al., 2019). CWSI varies between 0 and 1, where 0 corresponds to a lower limit indicating a potentially transpiring crop and 1 corresponds to the upper limit indicating a non-transpiring crop. CWSI has been used for detecting stress, scheduling irrigation and predicting yield in different crops across various agro-climates (Gontia and Tiwari, 2008, Anda, 2009, Akkuzu et al., 2013, González-Dugo et al., 2014, Bahmani et al., 2017, Anda et al., 2020).

The calculation of CWSI involves the use of upper baseline (Tcanopy-ub) and lower baseline (Tcanopy-lb) canopy temperatures. These baseline temperatures are generally obtained using either the theoretical approach (Jackson et al., 1981) or the empirical approach (Idso et al., 1981). In the former approach, the baseline temperatures are computed using the energy balance method based on net radiation, aerodynamic resistance, and canopy resistance values. Though, this approach precisely determines the CWSI, the parameters required are complex and infeasible in field conditions (Alghory and Yazar, 2019). In the latter approach, Tcanopy-ub and Tcanopy-lb are determined through a linear regression between Tcanopy – Tair and vapour pressure deficit for non-transpiring and potentially transpiring crops, respectively. A constant has been recommended to be used for estimating Tcanopy-ub (King and Shellie, 2018). This approach provides reliable estimates of CWSI and is relatively simpler in field applications.

It is clear from the above that the reliability of CWSI depends on the Tcanopy-ub and Tcanopy-lb values (Cohen et al., 2017). Although direct observation of these baseline temperatures would be ideal, it is practically difficult to determine the values for Tcanopy-lb and Tcanopy-ub while simultaneously measuring Tcanopy, as both involve unattainable or undesirable field moisture conditions (Pramanik et al., 2017, Kumar et al., 2020a). Alternate methods for estimating the baseline temperatures have been suggested viz. artificial wet and dry reference surfaces (Pou et al., 2014), well-watered and stressed treatments (Möller et al., 2007), statistical or bio-indicators approach (Alchanatis et al., 2010), numerical estimation through physical models (Jones, 1999), etc. However, their field application is limited due to complex parameters involved in the physical models and extensive maintenance of reference surfaces (Park, 2018; Jamshidi et al., 2021). Hence, the development of parsimonious models, which do not essentially require the baseline temperatures and can provide precise estimates of the CWSI, will improve the utilization of CWSI in crop stress management and irrigation scheduling.

Data-driven neural network modelling techniques have been successfully used to model complex non-linear algorithms to predict physical variables (Adeloye et al., 2012, Ramachandran et al., 2019). These techniques are found to yield reliable results without needing any physical or mathematical description of the model, even with minimum data (Maroufpoor et al., 2018, King et al., 2020). The most widely used technique is the supervised artificial neural network (ANN) which is a multi-layer perceptron network and used feed-forward back-propagation algorithm. ANN has been extensively applied to model numerous water resources phenomena (ASCE Task Committee on Application of Artificial Neural Networks in Hydrology, 2000, Campos de Oliveira et al., 2017, Roshni et al., 2020, Walls et al., 2020). However, ANN modelling and prediction performance significantly reduce when noise (outliers or missing values) is present in the data (Rustum, 2009). Such shortcomings can be effectively handled by using unsupervised algorithms such as the Self-organizing map (SOM), where the data are easily interpreted and understood and where the clustering helps in identifying the similarities in the data (Kohonen et al., 1996, Adeloye and Rustum, 2012). SOM has also been widely applied in water resources modelling and prediction purposes (Kalteh et al., 2008, Mwale et al., 2012, Rodríguez-Alarcón and Lozano, 2017, Chen et al., 2018, Brentan et al., 2018, Ohana-Levi et al., 2019).

Despite the proven potential of these techniques, their application in modelling the CWSI is very limited. King and Shellie, 2016, King and Shellie, 2018 reported on the application of ANN for predicting the Tcanopy-lb of wine grape cultivars using Tair, RH, solar radiation and wind speed. They found that the ANN models performed exceptionally well and provided reliable estimates of Tcanopy-lb. Kumar et al. (2021) evaluated the performance of ANN and SOM models in estimating the Tcanopy-lb for Indian mustard considering Tair and RH and found that the SOM estimates outperformed the ANN estimates. Kumar et al. (2020a) explored the potential of ANN and SOM for modelling the CWSI using easily attainable environmental data (Tair Tcanopy and RH). This approach did not require the baseline temperatures for predicting the CWSI. They found that SOM performed better than the ANN in predicting the CWSI. This was achieved by the powerful clustering capability of SOM which converts a high dimensional input into a low dimension output and extracts useful features (information) from the input vector. The authors have, however, not explored the SOM model and its structure in detail and suggested the need for investigation of the underlying features and clusters to provide a better understanding of the model.

The study aims to develop a SOM model for CWSI involving data collected in field experiments conducted on Indian mustard grown in a humid-subtropical location. The objectives are to:

  • (i)

    compute the CWSI empirically using the lower and upper baseline canopy temperatures;

  • (ii)

    train, test and evaluate the performance of the SOM model for predicting CWSI using easily attainable microclimatic variables;

  • (iii)

    explore and extract the cluster features of the developed SOM model;

  • (iv)

    analyze the performance of the SOM model for different levels of irrigation treatments.

The description of the study area, irrigation treatments, and modelling database is presented in Section 2. Additionally, the experimental and modelling methodology is given along with a brief description of SOM. In Section 3, the experimental and SOM modelling results along with discussion are presented. In Section 4, the conclusions are finally presented.

Section snippets

Study area and experimental site

The present study was conducted at the humid sub-tropical location of Hamirpur in the north Indian state of Himachal Pradesh. Field experiments involving canopy temperature and soil water measurements along with other microclimatic variables were performed in an agricultural experimental station (AES) located in the campus of the National Institute of Technology Hamirpur. The AES has an area of 0.055 acres and is located at 31° 42′ 40″ (latitude) and 76° 31′ 34″ (longitude) on an altitude of

Baseline computed CWSI

The CWSI was computed using the experimentally obtained baseline temperatures Tcanopy-lb and Tcanopy-ub. Fig. 8 shows the temporal variation of baseline computed CWSI for treatments I2 to I7 during the 2017 and 2018 crop periods. For each treatment, one-way ANOVA test statistics showed computed F < critical F (i.e., p > 0.05), indicating CWSI between the replications were not different. Therefore, the mean of three replications is used to represent the variation of baseline computed CWSI in

Conclusions

In this study, we presented the application of SOM to model and predict the CWSI. Field experiments on Indian mustard, involving different levels of irrigation treatments, were performed to generate varying degrees of CWSI. The training and testing of the SOM model utilized the data on noon-time measurements of the microclimatic variables Tair, Tcanopy and RH, recorded during the crop growth period. The baseline computed CWSI was used as the reference CWSI. The powerful visualization capability

CRediT authorship contribution statement

Navsal Kumar: Conceptualization, Software, Formal analysis, Investigation, Writing - original draft. Rabee Rustum: Methodology, Software, Validation, Writing - review & editing. Vijay Shankar: Supervision, Project administration, Writing - review & editing, Funding acquisition. Adebayo J. Adeloye: Funding acquisition, Visualization, Writing - review & editing.

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.

Acknowledgements

The support extended by the staff of the National Institute of Technology Hamirpur (India) and Heriot-Watt University (United Kingdom) is gratefully acknowledged.

Funding

The research is an outcome of the overseas visiting doctoral fellowship awarded to Navsal Kumar (Award No. ODF/2018/000374) by SERB, Department of Science & Technology (Govt. of India).

The study forms an integral component of the research project “Sustaining Himalayan Water Resources in a changing climate (SusHi-Wat)” supported by the NERC-MoES Newton-Bhabha Fund for Sustaining Water Resources for Food, Energy & Ecosystem Services in India (Award No. NE/N016394/1 (UK-NERC);

Data Availability Statement

Some data, models or code used during the study are available from the corresponding author by request.

Software Availability

The SOM Toolbox (version 2.1) for MATLAB is an open-source and publicly available to download from GITHUB (https://github.com/ilarinieminen/SOM-Toolbox).

References (55)

  • B.A. King et al.

    Evaluation of neural network modeling to predict non-water-stressed leaf temperature in wine grape for calculation of crop water stress index

    Agric. Water Manage.

    (2016)
  • B.A. King et al.

    Wine grape cultivar influence on the performance of models that predict the lower threshold canopy temperature of a water stress index

    Comput. Electron. Agric.

    (2018)
  • N. Kumar et al.

    Neural computing modelling of the crop water stress index

    Agric. Water Manage.

    (2020)
  • F.D. Mwale et al.

    Infilling of missing rainfall and streamflow data in the Shire River basin, Malawi–A self-organizing map approach

    Phys. Chem. Earth, Parts A/B/C

    (2012)
  • N. Ohana-Levi et al.

    A weighted multivariate spatial clustering model to determine irrigation management zones

    Comput. Electron. Agric.

    (2019)
  • A.H. Orta et al.

    Crop water stress index for watermelon

    Sci. Hortic.

    (2003)
  • Y. Osroosh et al.

    Comparison of irrigation automation algorithms for drip-irrigated apple trees

    Comput. Electron. Agric.

    (2016)
  • A. Pou et al.

    Validation of thermal indices for water status identification in grapevine

    Agric. Water Manag.

    (2014)
  • A. Ramachandran et al.

    Anaerobic digestion process modeling using Kohonen self-organising maps

    Heliyon

    (2019)
  • T. Vatanen et al.

    Self-organization and missing values in SOM and GTM

    Neurocomputing

    (2015)
  • M. Weiss et al.

    Remote sensing for agricultural applications: A meta-review

    Remote Sens. Environ.

    (2020)
  • A.J. Adeloye et al.

    Self-organising map rainfall-runoff multivariate modelling for runoff reconstruction in inadequately gauged basins

    Hydrol. Res.

    (2012)
  • E. Akkuzu et al.

    Determination of crop water stress index and irrigation timing on olive trees using a handheld infrared thermometer

    J. Irrig. Drain. Eng.

    (2013)
  • V. Alchanatis et al.

    Evaluation of different approaches for estimating and mapping crop water status in cotton with thermal imaging

    Precis. Agric.

    (2010)
  • A. Alghory et al.

    Evaluation of crop water stress index and leaf water potential for deficit irrigation management of sprinkler-irrigated wheat

    Irrig. Sci.

    (2019)
  • A. Anda

    Irrigation timing in maize by using the crop water stress index (CWSI)

    Cereal Res. Commun.

    (2009)
  • ASCE Task Committee on Application of Artificial Neural Networks in Hydrology. (2000). Artificial neural networks in...
  • Cited by (13)

    View all citing articles on Scopus
    View full text