New spectral model for estimating leaf area index based on gene expression programming

https://doi.org/10.1016/j.compeleceng.2020.106604Get rights and content

Abstract

The accurate evaluation of leaf area index (LAI) in paddy rice is of great importance in predicting the production and assessing food security. Traditional LAI estimation based on the PROSAIL model requires a large number of parameters. To solve this problem, this paper proposes a new estimation model between LAI and spectral reflectance for paddy rice canopies based on gene expression programming. And, we suggest a new spectral index for maximizing the response to LAI and weakening the sensitivity to chlorophyll concentration in the proposed LAI estimation model. Experimental results show that compared with existing indices, the proposed index (VI-R2) maximized the response to green LAI and was less sensitive to chlorophyll content variations, and in the validation with measured LAI values, VI-R2 shows the excellent fitting performance with coefficient of determination (r2) of 0.954. Meanwhile, based on a comparison of the estimated LAI values among existing indices using hyperspectral image, the results indirectly demonstrate that VI-R2 is usable for estimating LAI at the remote sensing pixel scale.

Introduction

Rice is one of the most important food crops in the world. Rice was the staple food of approximately 2.9 billion people living in 33 countries in 1993 [1]. In 2013, the total production of paddy rice around the world was approximately 745 million tons, which is only less than that of maize [2]. To predict the rice production in rice-growing regions and countries, we need to obtain information about rice growth status in a timely and accurate manner and take rapid and effective measures. Leaf area index (LAI) is the key factor that determines the photosynthetic capacity of a canopy and is important for the carbon build-up in the ecosystem, the productive power of the vegetation and the energy balance between soil, vegetation and atmosphere [3]. Estimating the leaf area index of rice can satisfy the requirements of monitoring rice growth and predicting production.

Currently, we can obtain LAI through ground-based direct measurements of LAI and indirect estimates. The former is costly and time consuming to obtain LAI values over large areas. Many researchers have evaluated LAI using the spectral data in remote sensing images due to its rapid update advantages [4], [5], [6], [7], [8], [9], [10]. Vegetation indices have been proposed based on the relationships between the vegetation reflectance and biophysical processes and characteristics. During the last decades, numerous studies have made efforts to develop and improve vegetation indices that enhance vegetation reflectance information and weaken the effects of exogenous factors [11]. Recently, some scholars use machine learning algorithms to study LAI prediction [12], [13], [14], [15], which provides important references for various artificial intelligence algorithms in the field of spectral index construction on LAI.

The early intrinsic vegetation index NDVI [16] was used widely, but it saturates with increasing green LAI [17]. The Renormalized Difference Vegetation Index or RDVI [18] was developed to linearize the relationship with vegetation biophysical variables (e.g., LAI), and the Modified Simple Ratio, or MSR [19], was proposed with the aim of enhancing the sensitivity to vegetation biophysical variables. To minimize the effects of soil background, some soil-line vegetation indices were introduced [20]. The effect of variations in chlorophyll concentrations on LAI estimates was more or less present in the available vegetation indices [17]. To enhance the sensitivity to the LAI of crops and minimize the effect of chlorophyll content changes, Haboudane et al. [17] developed the vegetation indices MCARI2 and MTVI2 based on the green peak, the near-infrared and the minimum reflectance in the red region.

A few vegetation indices available are proposed specifically for estimating the LAI in paddy rice canopies. Kimura et al. [21] proposed a vegetation index for LAI in paddy rice (VILAI) using the green band (550 nm) and NIR (near-infrared) band (980 nm), based on the concepts of MSAVI. Specially, rice is planted in paddies (shallow puddles) that are carefully controlled to ensure appropriate water depth and amount of algae growing on the water surface of the flooded paddies during early growth. This is often the case in real world situations [1], which is the greatest difference between paddy rice and other food crops in terms of their growth. Thus, for estimating LAI in rice, the effect of the algae on the water surface may be different from the effect of a soil background. In previous studies, the proposed vegetation index is always based on the concepts of previous indices and experience or statistical methods [22], [23]. This process is useful for developing a new better index to estimate biophysical or biochemical parameters of vegetation. However, the selection of bands and functional form of the new vegetation indices are dependent on previous vegetation indices. Perhaps another functional relationship or combination between spectral bands is not taken into account for estimating biophysical or biochemical parameters of vegetation by previous indices. This combination may enhance the sensitivity of the subject factor (e.g., green LAI of canopy) and weaken the sensitivity to interference factors.

Gene expression programming algorithm (GEP) is an evolutionary algorithm that was proposed by the Portuguese biologist Ferreira in 2001 [24], [25]. GEP uses a search and optimization technique to determine the most fit model from dataset through artificial evolution [26], [27]. The fittest model can help to find the underlying rules from dataset. Compared with traditional genetic algorithm (GA) [28] and genetic programming (GP) [29], the convergence speed of GEP is improved by 4–6 times [24]. Compared with the existing function fitting approaches based on regression analysis, the GEP does not need to set a function model to be discovered in advance [25], which greatly improves the objectivity of the function model obtained through mining. Leaf area index estimation is to find the function model between leaf area index and spectral features to judge the spectral features that affect leaf area index. Therefore, this paper proposes a new spectral model for estimating leaf area index based on gene expression programming. For canopy parameter estimation, the important spectral bands from more than a thousand narrow-band can be selected and the function of the canopy parameter (e.g., LAI) with respect to the spectral bands can be mined by GEP. Relative to the conventional regression analyses, an a priori functional form does not need to be specified by GEP, which performs symbolic regression. Compared to the methods based on empirical knowledge or conventional spectral indices, GEP can exploit the important spectral information or formula that may be neglected. The major contributions of our work are listed as follows:

  • To quantitatively analyze the relationship between canopy parameter estimation and canopy spectral reflectance, this paper proposes a new method on canopy parameter estimation between canopy parameters and canopy spectral reflectance based on gene expression programming.

  • On the basis of this, in this paper, we propose a new spectral index using this method to maximize the response to green LAI in paddy rice canopies and weakening the sensitivity to chlorophyll concentration. Experimental results show that for the new spectral index, the saturation level of the LAI estimation and the resistance to the chlorophyll effect were tested using the combined leaf and canopy radiative transfer PROSPECT + SAIL model.

  • Experimental results on real hyperspectral image and simulated data based on PROSAIL model showed that the proposed model in this paper outperformed the traditional statistical model in terms of r2, RMSE, spectral reflectance, intercept and slope.

Fig. 1 shows the flowchart of deriving the new spectral index for LAI estimation by GEP in this study. The leaf area index, only including green leaves, was defined as LAI in this study.

The remainder of this paper is organized as follows. Section 2 introduces experimental data sources and related methods. Section 3 analyzes and discusses the experimental results from chlorophyll effect, validation and LAI estimation. Finally, conclusions are given in Section 4.

Section snippets

Experimental site and design

This study area is located in a paddy rice field in Ma an township in Liuhe District, Nanjing, Jiangsu Province, China 3222’N, 11851’E. The geography of the study area is shown in Fig. 2. The experiment area is classified as a typical subtropical monsoon climate region, with a multi-year mean rainfall of 1100 mm and with a multi-year mean temperature of 15.4 C. The soil texture is sandy loam paddy soil with a pH of 6.8, organic matter content of 18.4 g/kg and total nitrogen (N) content of

Simulated chlorophyll effect and saturation level

With LAI variations in paddy rice, the range of variations in chlorophyll concentration also varies somewhat. Fig. 4 shows the simulated effect of variations in chlorophyll concentration on the relationship between vegetation index (existing vegetation indices and new spectral indices) values and simulated LAI for paddy rice using the PROSAIL model. Additionally, Fig. 4 compares the saturation level in the estimation of LAI among the vegetation indices.

VI-LAI is the vegetation index for paddy

Conclusions

This paper was designed to propose a new spectral index, VI-R2, for estimating the LAI in paddy rice using a new method based on gene expression programming algorithm (GEP). The major results of this paper are as follows.

The gene expression programming algorithm (GEP), an evolutionary algorithm, was applied to develop a new spectral index for estimating the LAI in paddy rice by constructing GEP mathematical models between the LAI and spectral reflectance of paddy rice canopies. The results

Declaration of Competing Interests

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.

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

CRediT authorship contribution statement

Lechan Yang: Conceptualization, Methodology, Software, Writing - original draft, Formal analysis. Song Deng: Data curation, Writing - review & editing, Funding acquisition. Zi Zhang: Resources, Investigation, Supervision.

Acknowledgments

We would like to thank the anonymous reviewers for their comments and constructive suggestions that have improved the paper. This work was supported by the National Natural Science Foundation of P. R. China (No. 61962006,51977113) and NUPTSF (No. NY219095).

References (44)

  • G. Rondeaux et al.

    Optimization of soil-adjusted vegetation indices

    Remote Sens Environ

    (1996)
  • R. Kimura et al.

    Relationships among the leaf area index, moisture availability, and spectral reflectance in an upland rice field

    Agric Water Manage

    (2004)
  • C. Daughtry et al.

    Estimating corn leaf chlorophyll concentration from leaf and canopy reflectance

    Remote Sens Environ

    (2000)
  • D. Haboudane et al.

    Integrated narrow-band vegetation indices for prediction of crop chlorophyll content for application to precision agriculture

    Remote Sens Environ

    (2002)
  • M.Z. Hashmi et al.

    Statistical downscaling of watershed precipitation using gene expression programming (GEP)

    Environ Modell Softw

    (2011)
  • H. Wang et al.

    Gene expression programming algorithms for optimization of water distribution networks

    Procedia Eng

    (2012)
  • R.S. Murray

    Advanced mathematics for engineers and scientists

    (1980)
  • W.A. Allen et al.

    Interaction of isotropic light with a compact plant leaf

    JOSA

    (1969)
  • W. Verhoef

    Light scattering by leaf layers with application to canopy reflectance modeling: the sail model

    Remote Sens Environ

    (1984)
  • D. Major et al.

    Accuracy and sensitivity analyses of sail model-predicted reflectance of maize

    Remote Sens Environ

    (1992)
  • M.M. Verstraete et al.

    Designing optimal spectral indexes for remote sensing applications

    IEEE Trans Geosci Remote Sens

    (1996)
  • N.H. Broge et al.

    Comparing prediction power and stability of broadband and hyperspectral vegetation indices for estimation of green leaf area index and canopy chlorophyll density

    Remote Sens Environ

    (2001)
  • Cited by (14)

    • A comparison of methods to estimate leaf area index using either crop-specific or generic proximal hyperspectral datasets

      2023, European Journal of Agronomy
      Citation Excerpt :

      Currently, three main types of approaches have been developed to estimate the crop LAI based on spectral information: 1) estimating LAI based on the simple statistical regression relationship between LAI and the spectral vegetation indices (Dong et al., 2019; Feng et al., 2019; Houborg and McCabe, 2018; Kira et al., 2016; Li et al., 2014; Liang et al., 2015; Mananze et al., 2018; Roosjen et al., 2018; Sun et al., 2020, 2018; Van der Meer et al., 2001; Vina et al., 2011; Yang et al., 2020; Zhang et al., 2021b; Zhao et al., 2007); 2) estimating LAI with machine learning regression methods, such as neural networks (Camacho et al., 2021; Chen et al., 2020; Liang et al., 2015; Pan et al., 2019; Verrelst et al., 2015a, 2015b), support vector regression (Camacho et al., 2021; Mananze et al., 2018; Panigrahi and Das, 2020; Verrelst et al., 2015a, 2012; Zhang et al., 2021a), or partial least squares regression (Li et al., 2014; Panigrahi and Das, 2020; Schiefer et al., 2021; Thorp et al., 2015; Verrelst et al., 2015a, 2015b; Wang et al., 2018; Zhang et al., 2021a); and 3) estimating LAI based on crop canopy radiative transfer models (Duan et al., 2014; Duveiller et al., 2011; Fang et al., 2003; Pan et al., 2019; Richter et al., 2011; Schiefer et al., 2021; Verrelst et al., 2015a; Wang et al., 2018; Weiss et al., 1999). Spectral vegetation indices are mathematical combinations of different spectral bands, and their main purpose is to enhance the information of interest about vegetation variables (e.g., LAI) and minimize the effects of soil, observation geometry, and other vegetation characteristics (Baret and Guyot, 1991; Campos-Taberner et al., 2016; Moulin and Guerif, 1999; Roosjen et al., 2018; Sun et al., 2020; Yang et al., 2020; Zhao et al., 2007). Simple statistical regression and machine learning regression methods have been widely used to estimate LAI of various vegetation species owing to their ease of application.

    • Estimation model of leaf nitrogen content based on GEP and leaf spectral reflectance

      2022, Computers and Electrical Engineering
      Citation Excerpt :

      However, it is impossible to estimate the actual nitrogen demand of crops more accurately based on human experience. Therefore, accurately estimating the nitrogen content of crop leaves is of great significance for monitoring and evaluating the quality of crop products [1]. In the existing study, it is difficult to accurately predict and estimate the leaf radiation transmission because that the nitrogen content of plant cannot be incorporated into the absorption and scattering of PROSPCT model.

    • Spontaneous facial expression database of learners’ academic emotions in online learning with hand occlusion

      2022, Computers and Electrical Engineering
      Citation Excerpt :

      Academic emotion can also be communicated to the outside world through learners’ facial expressions, which has been evidenced by a large number of relevant studies and daily teaching. Artificial intelligence and machine learning have been widely used in various fields [1–11], such as image retrieval [12], image captioning [13,14], semantic segmentation [15,16], facial expression recognition and so on. Some studies have explored the application of expression recognition technology in an online learning environment, and thus, preliminarily verified the application's effect.

    • Visual-based data exchange system for internal and external networks in physical isolation

      2021, Cognitive Robotics
      Citation Excerpt :

      Compared with NFC (near Field wireless communication technology), RFID (radio frequency technology) and other automatic identification technology, QR code also has the advantages of strong recognition ability, fast speed, anti-interference, fault tolerance and low cost. QR code is an essential part of our daily life, it can be used as the main carrier of information exchange, can also be used as a simple means of confidentiality[19–40].[ 19-40]

    View all citing articles on Scopus

    This paper is for CAEE special section on Artificial Intelligence and Computer Vision (VSI-aicv4). Reviews processed and recommended for publication to the Editor-in-Chief by Guest Editor CYJ Li.

    View full text