Comparative analysis of different uni- and multi-variate methods for estimation of vegetation water content using hyper-spectral measurements

https://doi.org/10.1016/j.jag.2013.04.004Get rights and content

Highlights

  • Comparison between most common univariate and multivariate techniques to be able to estimate canopy water content.

  • Utilizing hyperspectral data to better understand the effect of biochemical parameters of vegetation on their reflectance.

  • Estimation of canopy water content as an important canopy parameter which has not been much studied before.

Abstract

Assessment of vegetation water content is critical for monitoring vegetation condition, detecting plant water stress, assessing the risk of forest fires and evaluating water status for irrigation. The main objective of this study was to investigate the performance of various mono- and multi-variate statistical methods for estimating vegetation water content (VWC) from hyper-spectral data. Hyper-spectral data is influenced by multi-collinearity because of a large number of (independent) spectral bands being modeled by a small number of (dependent) biophysical variables. Therefore, some full spectrum methods that are known to be suitable for analyzing multi-collinear data set were chosen. Canopy spectral reflectance was obtained with a GER 3700 spectro-radiometer (400–2400 nm) in a laboratory setting and VWC was measured by calculating wet/dry weight difference per unit of ground area (g/m2) of each plant canopy (n = 95). Three multivariate statistical methods were applied to estimate VWC: (1) partial least square regression, (2) artificial neural network and (3) principal component regression. They were selected to minimize the problem related to multi-collinearity. For comparison, uni-variate techniques including narrow band ratio water index (RWI), normalized difference water index (NDWI), second soil adjusted vegetation index (SAVI2) and transferred soil adjusted vegetation index (TSAVI) were applied. For each type of vegetation index, all two-band combinations were evaluated to determine the best band combination. Validation of the methods was based on the cross validation procedure and using three statistical indicators: R2, RMSE and relative RMSE. The cross-validated results identified PLSR as the regression model providing the most accurate estimates of VWC among the various methods. The result revealed that this model is highly recommended for use with multi-collinear datasets (RCV2=0.94, RRMSECV = 0.23). Principal component regression exhibited the lowest accuracy among the multivariate models (RCV2=0.78, RRMSECV = 0.41).

Introduction

The quantification of biophysical and biochemical factors is a key element in vegetation monitoring and understanding terrestrial processes (Goetz et al., 1992). Remote sensing is regarded as a fast and non-destructive technique that has been widely used in quantifying biophysical (Cohen et al., 2003, Cho et al., 2007) and biochemical parameters (Curran et al., 2001, Hansen and Schjoerring, 2003, Darvishzadeh et al., 2008a, Darvishzadeh et al., 2008b) in different scales. Water content is one of the main properties of vegetation that can be investigated by remotly sensed data. Vegetation water status plays an important role in plant functioning, water and energy exchange with the atmosphere, as well as drought and fire risk (Penuelas et al., 1993, Penuelas et al., 1996). Vegetation water content (VWC) has been estimated in numerous studies (Clevers et al., 2010, Colombo et al., 2008, Sims and Gamon, 2003, Ceccato et al., 2001, Ceccato et al., 2002, Cheng et al., 2006, Danson and Bowyer, 2004). In the mentioned studies, different definitions were used for describing the VWC such as equivalent water thickness (EWT), fuel moisture content (FMC) gravity water content (GWC), and canopy water content (CWC). Remotely sensed VWC has been used to assess plant water status for irrigation purpose and was found helpful in crop estimation (Hunt and Rock, 1989, Penuelas et al., 1993, Jackson et al., 2004), retrieval of soil moisture (Jackson, 1993, Yilmaz et al., 2008), assessing vegetation conditions related to drought (Tucker, 1980, Claudio et al., 2006) and monitoring the risk of forest fires (Pyne et al., 1996, Maki et al., 2004).

Numerous research efforts focus vegetation indices from broad spectral bands for estimating VWC and other vegetation parameters (Chen et al., 2005, Sims and Gamon, 2003, Fensholt and Sandholt, 2003). One of the most important drawbacks of these multispectral products is the use of average information over broadband widths that results in the loss of critical information provided by narrow bands (Thenkabail et al., 2000). Hyper-spectral remote sensing (or imaging spectroscopy) enables the characterization of vegetation optical properties in many small, contiguous spectral bands within the visible, NIR and shortwave infrared regions. It has demonstrated great potential for an accurate estimation of vegetation water content (Clevers et al., 2010).

Previous studies have shown for example that narrow bands can be critical in providing essential information for quantifying the biophysical and biochemical characteristics of vegetation (Thenkabail et al., 2000, Broge and Leblanc, 2001, Darvishzadeh et al., 2008a, Clevers et al., 2010, Zhang et al., 2003). However, it is complex and time consuming to analyze a large number of contiguous narrow bands provided by hyper-spectral data. Therefore, when hyper-spectral data is available researchers often try to utilize methods that can summarize a large number of these multi-collinear narrow spectral bands. Approaches involving uni-variate and multi-variate techniques have been widely used to estimate vegetation parameters and crop yield (Sims and Gamon, 2003, Atzberger et al., 2004, Nguyen and Lee, 2006, Cho et al., 2007, Ye et al., 2007, Li et al., 2008, Darvishzadeh et al., 2008b). Uni-variate techniques based on spectralindices are still widely used to predict vegetation biophysical properties such as vegetation water content (Sims and Gamon, 2003, Jackson et al., 2004, Thenkabail et al., 2000, Maki et al., 2004, Chen et al., 2005). At most, indices use two or three bands for detecting relationships between spectral measurements and may be limited in terms of exploiting the rich information in narrow bands of hyper-spectral data. Alternatively, several studies have focused on multi-variate models, such as multiple linear regression and stepwise regression, which use several spectral wavelengths for estimating biophysical and biochemical properties (Curran et al., 2001). However, these methods – when used with hyper-spectral data – are likely to suffer from multi-collinearity (Curran, 1989, Grossman et al., 1996).

Multi-collinearity is a common problem inherent to hyperspectral dataset. It arises when one or more of the independent variables (narrow bands) are highly correlated with one or more other independent variables (Curran et al., 2001, Nguyen and Lee, 2006, Van der Meer and Jia, 2012). Multivariate statistical models such as partial least square regression (PLSR), artificial neural network (ANN) and principal component regression (PCR) address this problem (Wold et al., 2001, Atzberger et al., 2010). However, by using PCR and PLSR models, the effects of the multi-collinearity problem can only be reduced not completely removed (Da-Wen, 2010).

Partial least square regression and principal component regression are full-spectrum techniques that have been used for example to estimate vegetation parameters (Patel and Majumdar, 2010, Hansen and Schjoerring, 2003, Nguyen and Lee, 2006, Ryu et al., 2011), crop yield (Yang, 2011, Ye et al., 2007) and soil properties (Ramadan et al., 2005, Farifteh et al., 2007, Nocita et al., 2011, Vohland et al., 2011). However, there are few studies that have investigated the potential of PLSR for estimating VWC (Li et al., 2008).

Artificial neural networks have been widely used in remote sensing (Atkinson and Tatnall, 1997, Walthall et al., 2004, Ramadan et al., 2005, Noh et al., 2006, Liu et al., 2010). Many applications use this non-parametric method for land cover/land use classifications (Bruzzone et al., 1997, Townshend et al., 1991, Serpico et al., 1996, Weng, 2012). Additionally, non-linear relationships and non-Gaussian distributions of data can be modeled, soil and vegetation parameters, as well as crop yield, can be predicted (Ahmad et al., 2010, Yang et al., 2009, Ye et al., 2006, Farifteh et al., 2007). Hence, the potential of neural networks for VWC estimation needs to be evaluated.

With the present study we examine the utility of hyper-spectral measurements to estimate vegetation water content by applying different uni-variate and multi-variate statistical models. Three full spectrum methods involving PLSR, ANN and PCR were used and compared with narrow band vegetation indices. The suitability of each method was analyzed and compared in terms of the relative cross validated root mean square error (RRMSECV) and cross-validated coefficient of determination (RCV2). To fully control the measurement conditions, the study was based on laboratory canopy spectral measurements using four types of plants with different leaf size and shape.

Section snippets

Laboratory data collection

Four different plant species with different leaf shapes and sizes were selected for testing the utility of different methods for non-destructively estimating VWC. Fig. 1(a) illustrates the plant species and their variability in leaf size and shape. A total of 24 plants were used for the study, 6 plants per species. As different factors affect on canopy spectra, we generated variability within each species by inducing variation in canopy structure (i.e. LAI; see Section 2.2), canopy water

VWC estimation and hyper-spectral vegetation indices

Regions within the 2D-correlogram with coefficient of determination between VWC and narrow band indices higher than 75% are highlighted in Fig. 4. Each meeting point in the highlighted area of this plot corresponds to R2 value (R2 > 0.75) from the vegetation index made of the reflectance values indicated on the corresponding X and Y axes. The 2D-correlogram shows that the SWIR region includes many band combinations with high R2 values.

The best band position for all indices and R2 values are

Discussion

Our research demonstrated that hyper-spectral measurements, can successfully predict vegetation water content across four structurally different plant types with different LAI and soil backgrounds when combined with multivariate linear and non-linear statistical methods. Within a laboratory setting, the vegetation water content estimation in this study involved two statistical techniques: uni- and multi-variate statistical methods. We have attempted to highlight the ability of multi-variate

Conclusion

Our results demonstrate the potential of combining hyper-spectral data and multi-variate linear/non-linear methods to predict vegetation water content. However, despite their simplicity, uni-variate statistical methods based on optimized narrow band indices also showed satisfactory performance. Combined with ANN, optimized narrow band indices yielded even better results, because non-linearities are taken into account.

Comparing all methods used in this study, the highest correlation (R2 = 0.94)

References (74)

  • H.C. Claudio et al.

    Monitoring drought effects on vegetation water contentand fluxes in chaparral with the 970 nm water band index

  • J. Clevers et al.

    Estimating canopy water content using hyperspectral remote sensing data

    International Journal of Applied Earth Observation and Geoinformation

    (2010)
  • W. Cohen et al.

    An improved strategy for regression of biophysical variables and Landsat ETM+ data

    Remote Sensing of Environment

    (2003)
  • R. Colombo et al.

    Estimation of leaf and canopy water content in poplar plantations by means of hyperspectral indices and inverse modeling

    Remote Sensing of Environment

    (2008)
  • P.J. Curran

    Remote sensing of foliar chemistry

    Remote Sensing of Environment

    (1989)
  • P. Curran et al.

    Estimating the foliar biochemical concentration of leaves whit reflectance spectrometry. Testing the Kokaly and Clark methodologies

    Remote Sensing of Environment

    (2001)
  • F.M. Danson et al.

    Estimating live fuel moisture content from remotely sensed reflectance

    Remote Sensing of Environment

    (2004)
  • R. Darvishzadeh et al.

    Estimation of vegetation LAI from hyperspectral reflectance data: Effects of soil type and plant architecture

    International Journal of Applied Earth Observation and Geoinformation

    (2008)
  • R. Darvishzadeh et al.

    LAI and chlorophyll estimation for a heterogeneous grassland using hyperspectral measurements

    ISPRS Journal of Photogrammetry and Remote Sensing

    (2008)
  • J. Farifteh et al.

    Quantitative analysis of salt-affected soil reflectance spectra: a comparison of two adaptive methods (PLSR and.ANN)

    Remote Sensing of Environment

    (2007)
  • R. Fensholt et al.

    Derivation of a shortwave infrared water stress index from MODIS near- and shortwave infrared data in a semiarid environment

    Remote Sensing of Environment

    (2003)
  • B.-C. Gao

    NDWI—a normalized difference water index for remote sensing of vegetation liquid water from space

    Remote Sensing of Environment

    (1996)
  • Y.L. Grossman et al.

    Critique of stepwise multiple linear regression for the extraction of leaf biochemistry information from leaf reflectance data

    Remote Sensing of Environment

    (1996)
  • P. Hansen et al.

    Reflectance measurement of canopy biomass and nitrogen status in wheat crops using normalized difference vegetation indices and partial least squares regression

    Remote Sensing of Environment

    (2003)
  • T.J. Jackson et al.

    Vegetation water content mapping using Landsat data derived normalized difference water index for corn and soybeans

    Remote Sensing of Environment

    (2004)
  • L. Li et al.

    Retrieval of vegetation equivalent water thickness from reflectance using genetic algorithm (GA)-partial least squares (PLS) regression

    Advances in Space Research

    (2008)
  • M. Liu et al.

    Neural-network model for estimating leaf chlorophyll concentration in rice under stress from heavy metals using four spectral indices

    Biosystems Engineering

    (2010)
  • M. Maki et al.

    Estimation of water status to monitor the risk of forest fires by using remotely sensed data

    Remote Sensing of Environment

    (2004)
  • H.T. Nguyen et al.

    Assessment of rice leaf growth and nitrogen status by hyperspectral canopy reflectance and partial least square regression

    European Journal of Agronomy

    (2006)
  • M. Nocita et al.

    Predictions of soil surface and topsoil organic carbon content through the use of laboratory and field spectroscopy in the Albany Thicket Biome of Eastern Cape Province of South Africa

    Geoderma

    (2011)
  • H. Noh et al.

    A neural network model of maize crop nitrogen stress assessment for a multi-spectral imaging sensor

    Biosystems Engineering

    (2006)
  • Z. Ramadan et al.

    Application of PLS and Back-Propagation Neural Networks for the estimation of soil properties

    Chemometrics and Intelligent Laboratory Systems

    (2005)
  • C. Ryu et al.

    Multivariate analysis of nitrogen content for rice at the heading stage using reflectance of airborne hyperspectral remotesensing

    Field Crops Research

    (2011)
  • S. Serpico et al.

    An experimental comparison of neural and statistical non-parametric algorithms for supervised classification of remote-sensing images

    Pattern Recognition Letters

    (1996)
  • D.A. Sims et al.

    Estimation of vegetation water content and photosynthetic tissue area from spectral reflectance: a comparison of indices based on liquid water and chlorophyll absorption features

    Remote Sensing of Environment

    (2003)
  • P. Thenkabail et al.

    Hyperspectral vegetation and their relationships with agricultural crop characterstics

    Remote Sensing of Enviroment

    (2000)
  • J. Townshend et al.

    Global land cover classification by remote sensing: present capabilities and future possibilities

    Remote Sensing of Environment

    (1991)
  • Cited by (0)

    View full text