Mapping flooding regimes in Camargue wetlands using seasonal multispectral data

Remote Sensing of Environment 138 (2013) 165–171

Contents lists available at SciVerse ScienceDirect

Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse

Mapping flooding regimes in Camargue wetlands using seasonal multispectral data Aurélie Davranche a,⁎, Brigitte Poulin b, Gaëtan Lefebvre b a b

University of Angers, LETG Angers-LEESA, UMR CNRS 6554, 2 Boulevard Lavoisier, 49045 Angers Cedex, France Tour du Valat Research Center, Le Sambuc, 13200 Arles, France

a r t i c l e

i n f o

Article history: Received 24 February 2013 Received in revised form 17 July 2013 Accepted 19 July 2013 Available online xxxx Keywords: Classification tree Camargue Multispectral indices Regression models SPOT-5 Water levels Wetland monitoring

a b s t r a c t Reflectance data from multiseasonal SPOT-5 imagery was combined with monthly measures of water levels collected in the Rhône river delta (Camargue) in 2005 and 2006. Classification tree and regression models using monthly values of 17 multispectral indices and 4 bands, as well as their seasonal variations, were used for predicting the presence and levels of water, independently of vegetation type and density in shallow marshes. Accuracy of the classification model was estimated by cross-validation and by calculating the percentage of correctly classified pixels on the resulting maps using an independent sampling. Goodness-of-fit of the regression model was assessed by calculating the coefficient of correlation between predicted and observed values. Predictive accuracy of both models was estimated by calculating NRMSE for the independent validation sample. Regression model robustness was also tested using Scheffé post-hoc analyses on the residuals. Biophysical parameters of Camargue marsh vegetation were used to interpret misclassifications and model deficiency. Both models were composed of a single variable consisting of a multispectral index using the mid-infrared band. The resulting classification tree provided a cross-validation accuracy of 76% and a map validation accuracy of 83%. With an R = 0.5, the regression model predicted water level with a 6-cm precision up to 20 cm of water depth. For both approaches, the predictive power of model was most affected by close canopy. This study highlights the usefulness of data mining for long-term monitoring of wetland hydrology based on multispectral indices using the mid-infrared band. © 2013 Elsevier Inc. All rights reserved.

1. Introduction In the Camargue (Rhône river delta) water levels and salinity of wetlands are closely tied to human activities (Aznar, Dervieux, & Grillas, 2003). Development of rice farming over 60 years ago has translated into the building of a hydraulic network to introduce large volumes of freshwater from the Rhône River into the delta for field irrigation. Overtime, this ‘new’ water resource has become increasingly used for other human activities such as nature conservation, ecotourism, agriculture, waterfowl hunting, and fishing, the original unpredictable variations in water level and salinity being increasingly replaced by predictable permanent freshwater marshes. As a result, the species composition of aquatic habitats has changed from diversified Mediterranean to monospecific continental-type communities, resulting in biodiversity loss (Tamisier & Grillas, 1994). Robust monitoring tools are thus necessary to quantify flooding duration of various types of wetlands with the multiple objectives of monitoring biodiversity, conducting prospective and companion modeling related to climate changes (Mathevet et al., 2007), and elaborating a global delta management framework complying with European and national policies (Dervieux, 2005). ⁎ Corresponding author. Tel.: +33 241735049. E-mail address: [email protected] (A. Davranche). 0034-4257/$ – see front matter © 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.rse.2013.07.015

Multi-spectral data have been used for flood delineation with the assumption that water has very low reflectance in the near-infrared portion of the spectra, in contrast to other land features (Smith, 1997). However, flooded areas are frequently underestimated in the presence of emergent or submerged vegetation, typical of shallow wetlands (Smith, 1997). Radar remote sensing associated with multispectral satellite data has been used for modest classification schemes or to improve flooded wetland mapping, especially in areas with persistent cloud cover (Smith, 1997; Taft, Haig, & Kiilsgaard, 2004; Töyrä, Pietroniro, Martz, & Prowse, 2002). However, performance of active sensors is reduced by the presence of wind and vegetation (Sandoz, Chauvelon, Pichaud, & Buckwell, 2003; Smith, 1997), while radar imagery is affected by heavy rains and speckle, as well as turbidity changes (Delmeire, 1997). Horritt and Mason (2001) demonstrated that short wavelength radar SARs cannot detect water under vegetation, but that L-band and P-band SARs can penetrate vegetation canopy (Hess, Melack, Novo, Barbosa, & Gastil, 2003). However, long SAR wavelengths are not easily available, which represents a major constraint for repeated surveys (Aduah, Maathuis, & Hussin, 2007). These data are also more costly to acquire and more restrictive in terms of programming services compared to multispectral satellites that make feasible regular follow-ups by avoiding days with cloud cover (Davranche, Poulin, & Lefebvre, 2010; Jensen, Narumalani, Weatherbee, & Mackay,

166

A. Davranche et al. / Remote Sensing of Environment 138 (2013) 165–171

1993). Water level changes can be measured with InSAR, but the application works in wetlands covered by woody or herbaceous emergent vegetation only (Wdowinski et al., 2008). Geostat and TOPEX/POSEIDON NASA altimeters can be used to determine stage variations, but for water bodies greater than 1 km in width only (Al-Khudhairy et al., 2002). Non-parametric classifiers and generalized regression models using multispectral data have been successfully used recently for wetland monitoring based on their vegetation attributes (Davranche, Lefebvre, & Poulin, 2009; Poulin, Davranche, & Lefebvre, 2010). Extending these new perspectives to water management requires to take the following into account: the mid-infrared band is sensitive to soil moisture under the vegetation canopy (Baret, Guyot, Begue, Maur, & Podaire, 1988); the photosynthetic production of chlorophyll can be reduced with water decrease (Pettigrew, 2004; Souza, Machado, Silav, Lagoa, & Silveira, 2004); a situation of water stress can influence reflectance (Clay, Kim, Chang, Clay, & Dalsted, 2006); and several multispectral indices have proved to be useful for studying hydrological (Gao, 1996; Gond, Bartholome, Ouattara, Nonguierma, & Bado, 2004; Hanqiu, 2006; McFeeters, 1996) and biophysical (Bao, Liu, & Wang, 2008; Poulin et al., 2010; Vescovo & Gianelle, 2008) parameters. We could then hypothesize that detection of water through the vegetation might be improved when taking into account free water under the canopy, water content of the plant and photosynthetic activity altogether. This study aims at assessing the potential of seasonal series of multispectral data for mapping flooded areas and water levels in shallow wetlands regardless of vegetation types and density. For that purpose, we used classification trees and

generalized regression models, respectively, applied to a seasonal series of SPOT-5 scenes covering the Camargue.

2. Method 2.1. Study area The Camargue is a polderized floodplain of 145 300 ha that rarely exceeds an altitude of 5 m (Fig. 1). Its origin, both from the river and the sea, has produced landscapes marked by a horizontal gradient of water and salinity. With mean annual values of 600 mm of rainfall and 1400 mm of water evaporation, the Camargue is characterized by water deficits year round, especially in summer (Chauvelon, Tournoud, & Sandoz, 2003). These physical characteristics define the biotic components of the original ecosystems of Camargue, marked by a spatial and temporal unpredictable variability in their dynamics. These landscapes are no longer influenced by the geomorphological and climatic constraints only, being largely influenced by management of water levels. Agricultural activities, originally limited to alluvial deposits, have increased in areas due to rice farming that is also used in rotation with other crops to reduce soil salinity. The large private estates, resulting from loss of crops, are typically in lower areas subject to high salinity and frequent flooding. Most of these depressions have been embanked and are currently managed with more or less permanent water for waterfowl hunting (Dervieux, 2005). The Camargue is hence characterized by numerous fragmented wetlands varying in size, connectivity and hydrological regime.

Fig. 1. Flooding duration in Camargue according to the classification tree model for water presence in March, May, June, July, September and December 2005.

A. Davranche et al. / Remote Sensing of Environment 138 (2013) 165–171

167

2.2. Data acquisition

2.3. Statistical modeling

2.2.1. Space-based data Seasonal time series of SPOT-5 images (SPOT/Programme ISIS, Copyright CNES) were acquired in 2004–2005 and 2005–2006 in late December, March, May, June, July and September (August and October in 2006). These dates had been selected based on vegetation phenology and seasonal water management of the targeted wetlands described below. A single SPOT image covers the whole study area (60 × 60 km). SPOT-5 scenes have 10-m pixel resolution and four spectral bands: B1 (green: 0.50 to 0.59 μm), B2 (red: 0.61 to 0.68 μm), B3 (near-infrared NIR: 0.79 to 0.89 μm) and B4 (shortwave-infrared SWIR: 1.58 to 1.75 μm). SPOT scenes were acquired with radiometric correction of the distortions due to differences in the sensitivity of the elementary detectors of the viewing instrument, that is the preprocessing level called 1A (SPOT Image, 2008).

For modeling the presence or absence of water in Camargue marshes, a dichotomous partitioning (Breiman, Friedman, Olshen, & Stone, 1984) was performed with the Rpart (Recursive PARTitioning, Therneau & Atkinson, 1997) package in the R software. This method uses the cost complexity parameter (cp) for pruning. As described in Davranche et al. (2010), we used the cross-validation procedure called CV-1SE (Esposito, Malerba, Semeraro, & Tamma, 1999) for pruning with 10 subsets as well as iterative runs of the algorithm (Breiman et al., 1984) for the selection of the cp and, the prior parameter for imbalanced samples. Modeling of the water levels was performed with the Generalized Regression Model method using a forwardstepwise procedure in Statistica version 8.0 (StatSoft Inc.)

2.2.2. Ground data Selection of study plots resulted in a compromise between admittance, accessibility, and getting a representative sample of the Camargue wetlands based on aerial photographs and videos collected during survey flights. The number of plots sampled was further limited by the relatively short period of optimal plant growth. Field sampling was carried out at 108 plots (Fig. 1) located in seasonal or permanent shallow marshes either covered with tall (common reed Phragmites australis) or mid-size (club-rush Bolboschoenus maritimus) helophytes, or with submerged macrophytes (Potamogeton pectinatus, Potamogeton pusillus, Ruppia maritima, Myriophillum spicatum, Characea). Each plot was situated in a different hydrological unit at least 70 m from the marsh edge (see Davranche et al., 2010 for further details) and geolocated with a GPS (Holux GR-230XX). Vegetation sampling design was inspired from previous works on plant and bird ecology from these marshes (Poulin, Lefebvre, & Mathevet, 2005; Poulin, Lefebvre, & Mauchamp, 2002) adapted to the size and shape of SPOT pixels. For each plot, water levels, vegetation cover and floristic composition were estimated along two diagonals crossing a 20 × 20 m plot (4 pixels) between May and July, depending upon vegetation development. Water levels were systematically measured using a rule every 4 m along each diagonal and in the center of the plot (N = 17). Should part of the marsh by dry, level of underground water was estimated by digging a small hole until reaching the water table (negative value). When dryness was through several soil horizons, a value of 0 was attributed. Most of these marshes were equipped with a piezometer (200-cm PVC tube buried 50 cm under the ground surface at the edge of the marsh) or a rule (located in the deepest area of the marsh) monitored twice monthly. The latter were also monitored at the time of the plot survey in order to obtain a bi-monthly follow-up of the water level at the plot. When the date of image acquisition and of water monitoring differed by several days, water data were calibrated assuming linear change overtime. Common reed was the densest and tallest vegetation found in the marsh. Density of green and dry reed stems, their height and diameter were measured within four quadrats of 50 × 50 cm per plot located at 7 m from the center of the plot in each cardinal direction on both years. The ground surface covered by dry reeds (estimated by 3.14 × (mean stem diameter / 2)2 multiplied by the number of stems), as well as the number of leaves, their length, width and area were also measured in 2005. In club-rush habitat, height of stems and cover rate of the emerged and submerged species were estimated in 2005. In aquatic beds, we estimated the percent cover of the vegetation, the dominant plant species and the proportion of submerged plants showing at the water surface in 2005. For all habitat types, homogeneity of general plant cover was coded semiquantitatively (from 1 to 4), water salinity was measured with a conductivity meter, and water transparency was estimated using a Secchi disk.

SPOT-5 scenes were radiometrically corrected using the 6S atmospheric code (Davranche et al., 2009) developed by Vermote, Tanre, Deuze, Herman, and Morcrette (1997), and projected to Lambert conformal conic projection datum NTF (Nouvelle Triangulation Française) using a second-order transformation and nearest-neighbor resampling (RMSE b 1 pixel). The scenes were georeferenced to a topographic map at 1:25 000 scale. We extracted the mean reflectance value for each sampling plot from each band of each scene using the ‘Spatial Analyst’ of ArcGIS version 9.2 (Environmental Systems Research Institute, Meudon, France). Using these data, we further calculated for each plot the most common multispectral indices that we adapted to SPOT bands (Table 1). The MIFW was inspired by the IFW index used by Adell and Puech (2003) to spatially analyze hunting activity in Camargue marshes. These authors showed that the difference between the near-infrared and the green bands (IFW) of a LANDSAT TM image covering the Camargue in July 1999, had negative and low values for open water, but positive and high values for areas with emergent plants, permitting to classify pixels where

2.4. Image processing for the descriptive variables

Table 1 Multispectral and multitemporal indices used in this study. Indices

Formula

References

SR — simple ratio

B2 / B3

VI — vegetation index

B3 / B2

DVI — differential vegetation index MSI — moisture stress index NDVI — normalized difference vegetation index SAVI — soil adjusted vegetation index OSAVI — optimized SAVI

B3 − B2

Pearson and Miller (1972) Lillesand and Kiefer (1987) Richardson and Everitt (1992) Hunt and Rock (1989) Rouse, Haas, Schell, and Deering (1973) Huete (1988)

NDWI — normalized difference water index NDWIF — normalized difference water index of McFeeters MNDWI — modified normalized difference water index DVW — difference between vegetation and water IFW — index of free water MIFW — modified index of free water WII — water impoundment index MWII — modified water impoundment index WI — water index MWI — modified water index

B4 / B3 (B3 − B2) / (B3 + B2) 1.5 ∗ (B3 − B2) / (B3 + B2 + 0.5) (B3 − B2) / (B3 + B2 + 0.16) (B3 − B4) / (B3 + B4) (B1 − B3) / (B1 + B3) (B1 − B4) / (B1 + B4) NDVI − NDWI

Rondeaux, Steven, and Baret (1996) Gao (1996) McFeeters (1996) Hanqiu (2006) Gond et al. (2004)

B3 − B1 B4 − B1

Adell and Puech (2003) This study

B32 / B2

Caillaud et al. (1991)

2

This study

2

This study This study

B4 / B2 B3 / B1 B42 / B1

168

A. Davranche et al. / Remote Sensing of Environment 138 (2013) 165–171

the water was dominant. The MIFW associates the mid-infrared band (B4) of SPOT 5 with its green band (B1), accentuating their difference in reflectance values for water/moist surfaces. The MWII index is a modified version of the WII index used by Caillaud, Guillaumont, and Manaud (1991) to monitor coastal salt marshes in western France. The WII index was based on the lowest reflectance values of water surface in the near-infrared band of SPOT 1. It was shown to discriminate 5 cm or less of water from moist soil and could classify water depths varying from 1 to 60 cm with 73% of explained variance and residuals explained by different water colors. Inspired by this index, the MWII integrates the mid-infrared band of SPOT 5, which is sensitive to soil moisture under the vegetation canopy as shown by Baret et al. (1988). Capitalizing on this knowledge, we further calculated an index that we called WI and its modified version, the MWI both using the highest reflectance value of water in the green channel and the lowest values in the infrared bands. The term “modified” was added to the name of indices originally calculated with a near-infrared band.

2.5. The dependent variable The minimum and maximum levels among the eight water measurements taken on each diagonal from the field plot were extracted and grouped into the following categories: – Minimum and maximum ≤ 0 cm: Class 1 for totally dry plots – Minimum and maximum N 0 cm: Class 2 for totally wet plots – Minimum b 0 and maximum N 0 cm: Class 3 for partially flooded plots. Both successive years of sampling were grouped (2005 and 2006) in order to maximize the surveyed months and to take into account interannual variability in weather and/or water management. We further selected data belonging to classes 1 and 2 to detect dry and flooded areas, respectively. Although the third class 3 is interesting from an ecological point of view, it was discarded to reduce misclassification errors related to mixed pixels. For each habitat type and each scene, data from classes 1 and 2 were equally split into a training and validation sample. In the case of odd number, priority was given to the training sample. The training sample consisted of 72 plots of dry marsh and 215 plots of wet marshes, and the validation sample of 61 plots of dry marshes and 204 plots of wet marshes. Generalized regression modeling was used to predict the mean water level calculated from the sampling plot data.

Table 2 Error matrix for the classification of wet and dry marshes. Classification data

Producer's Error of User's Error of accuracy omission accuracy commission

Dry Wet Total Reference Dry 42 data Wet 25 Total 67

19 179 198

61 204 265

69 88

31 12

63 90

37 10

2.7. Interpretation of misclassifications and model power The binary response (0/1) for misclassified and well-classified plots in both years was confronted to structural parameters of reed or submerged macrophytes considered individually, using the likelihood ratio test (Davranche et al., 2010; Sokal & Rohlf, 1995) for model significance. For reed beds, a variable “year” was included as a potential parameter of misclassification. Then, a standardized residual analysis provided the response thresholds affecting the classification. Parameters influencing model accuracy from the GRM analysis were considered individually using a likelihood ratio test on the residuals. The Holm–Bonferroni multiple testing was applied. Then, NRMSE was calculated for each parameter class to highlight the values that affected model prediction. We further tested the three water classes used for the dependent variable (see Section 2.5). Scheffé post-hoc analyses were performed on the residuals to test model robustness according to water level range. For this test, all negative values were set to zero and positive values grouped into 5-cm interval classes. Holm–Bonferroni corrections for multiple tests were applied when relevant (Holm, 1979). 3. Results 3.1. Classification tree method The resulting tree was achieved with a prior parameter adjusted to 0.50 for both classes. It provided a cross-validation accuracy of 76% with 90% for wet marshes and 74% for dry marshes. The resulting formula was MIFW ≤ 0.1152. Mapping validation provided an overall accuracy of 83%, with 88% for wet marshes and 69% for dry marshes (Table 2). In reed beds, misclassifications were related to high and big green stems having long leaves, and low dry stem density (Table 3). None of the parameters collected in the club-rush and aquatic beds habitat could explain misclassifications. 3.2. Regression model

2.6. Validation The equations issued from the resulting decision tree and regression model were applied to SPOT-5 scenes of 2005 and 2006 for model validation. The raster calculator (Spatial Analyst) of ArcGIS was used to create the maps. The application of the classification tree resulted in binary maps where 1 encoded for wet marshes and 0 for dry marshes. The regression model produced a map with a gradient of colors (continuous values). Using the zonal statistics tool (Spatial Analyst) of ArcGIS, we extracted values 1 and 0 for water presence and absence, and water level values for each plot of the validation sampling. An error matrix was calculated using the validation pixel values (0 and 1) extracted from each monthly map resulting from the classification tree method (Table 2). Goodness-of-fit of the regression model was assessed by calculating the coefficient of correlation (R) and the normalized root-mean-square error (NRMSE) between the predicted and observed values of the training sample. Its predictive accuracy was assessed by calculating R and NRMSE between the predicted and observed values of the validation sample.

The model that best predicted water levels used a single multispectral index, providing an R coefficient of 0.5 (P ≤ 0.001) for both the training and validation samples. Model formula was: 22.3522034– 36.114709 ∗ MWII. The NRMSE was 18% for the training sample and 23% for the validation sample. Mean predicted and observed values of

Table 3 Effects and values of the parameters influencing the classification of flooded areas in Camargue reed beds. Parameters

Effects on classification

Values of parameters

P

Height (cm) Length of leaves (cm) Diameter of green stems (mm) Number of dry stems

− + − −

N300 20 b x ≤ 25 N40 6.5 b x ≤ 7

0.01 0.05 0.05 0.01

−

0 b x ≤ 100

0.01

A. Davranche et al. / Remote Sensing of Environment 138 (2013) 165–171

the validation samples differed significantly (t = 14.4, P ≤ 0.001), suggesting that field calibration might improve the result. The model provided a maximum threshold prediction at about 22 cm of water level. The model performance was better in 2006 than 2005. Water levels were better predicted in September, in reed beds, and for the class of water level coded 2. In contrast to other habitats, water levels in reed beds was better predicted in 2005 than 2006. Thick green stems having long leaves and low cover of apparent soil negatively influenced the prediction power of the model (Table 4). For aquatic beds, model accuracy was positively influenced by water transparency (Table 4). Water levels were better predicted in club-rush beds dominated by Bolboschoenus maritimus, a shorter plant (30–70 cm tall) than Scirpus littoralis (115–120 cm). Water levels were also less accurately predicted in sites presenting an important cover rate of submerged species, as well as for sites with 20–40% of apparent soil. Dry marshes were systematically predicted with a 2 cm precision, and wet sites within a 6-cm range (Sheffé post-hoc test, P = 0).

Table 4 Values of parameters affecting model power for the prediction of water levels. Habitats

Parameters

P

All types

Year

0

Month

Habitats

Class of water levels Reed beds

4. Discussion

Year Green stem diameter (mm)

4.1. The selected remote sensing variables The use of IFW (Adell & Puech, 2003) was based on a threshold, hence better adapted to categorical variables (e.g. presence/absence), whereas the WII (Caillaud et al., 1991) performed better with continuous environmental variables such as water levels. This might explain the selection of MIFW in the resulted tree and of MWII in the regression model analyses of this study. Although the NDWI (Gao, 1996) was developed to monitor water content in vegetation, Clay et al. (2006) showed that for an advanced development stage of corn species, it was not correlated to plant water stress, suggesting that water detection could be limited by plants with dense canopies. It could partly explain the selection of an index combining mid-infrared and green/red bands rather than one based on infrared spectra portions only. Indices integrating the mid-infrared band offer an additional advantage highlighted by Hanqiu (2006) when using the MNDWI inspired by the NDWIF (NDWI of McFeeters, 1996): they are less influenced by environmental noises. This characteristic is particularly interesting for coastal salt marshes that have high reflectance contrasts because of their contiguous, yet restricted areas (Caillaud et al., 1991).

Leave length (cm)

Cover of apparent soil (%)

Aquatic beds

Water transparency (cm)

Club rush Dominant species beds Height of emergent species (cm)

Cover of submerged vegetation (%)

4.2. Performance of the models No year effect could explain misclassifications, suggesting that the model from the classification tree was robust. The misclassifications observed in reed beds with fully grown and vigorous dense reed stems indicate that our method has still limitations for detecting water under close canopy, as observed in previous studies (Smith, 1997; Töyrä et al., 2002). The variable “year” significantly affected the regression model accuracy for all types of habitats. However, the NRMSE differed only by 1% between both years, suggesting that the robustness of the model might not be the reason. Only the 2005 sample included null instead of negative values for some very dry marshes, which could explain the year effect, since a lower correlation is obtained when the 2006 negative values are set to 0 (R2 = 0.27, P = 0). September provided the best model accuracy for all habitat types. This period is characterized by relatively high water levels, due both to seasonal rainfall and water control. This resulted in low midinfrared values (Davranche, 2008), reinforcing the spectral response characteristic of deeper waters with a MWII tending to 0. Water levels in reed beds are significantly better predicted compared to other habitats (Table 4). The higher overall accuracy could be related to the higher number of observations in reed beds, offering a

169

Cover of apparent soil (%)

Parameter values

2005 2006 0.05 March May June July August September October December 0 Aquatic beds Reed beds Club-rush beds 0.01 1 2 3 0 2005 2006 0.001 3 b x ≤ 3.5 3.5 b x ≤ 4 4 b x ≤ 4.5 4.5 b x ≤ 5 5 b x ≤ 5.5 5.5 b x ≤ 6 6bx 0 20 b x ≤ 25 25 b x ≤ 30 30 b x ≤ 35 35 b x ≤ 40 40 b x ≤ 50 0 x≤0 0 b x ≤ 20 20 b x ≤ 40 40 b x 0 0 b x ≤ 25 25 b x ≤ 50 50 b x ≤ 75 0.01 Scirpus littoralis Bolboschoenus maritimus 0.01 25 b x ≤ 50 50 b x ≤ 75 75 b x ≤ 100 100 b x ≤ 125 125 b x ≤ 150 0.01 0 0 b x ≤ 20 20 b x ≤ 40 40 b x ≤ 60 0.01 0 0 b x ≤ 20 20 b x ≤ 40 40 b x ≤ 60

Number of NRMSE observations (%) 318 244 75 97 100 94 32 48 29 87 108 382 62 140 422 191 197 185 22 51 93 81 27 38 17 13 52 93 60 21 33 152 26 28 27 25 44 20 42

9 8 12 14 14 12 11 9 10 10 14 6 14 39 25 34 18 19 23 20 20 22 23 32 84 26 21 20 23 35 39 22 22 24 32 25 20 30 17

13 29 0 8 12 13 44 0 5 20 12 5 25

27 9 61 26 27 21 50 23 26 50 20

better training basis for model building. Also, the model worked better in 2005 than 2006, again with only 1% difference of NRMSE. Spring rainfall differed between both years (664 mm in 2005 vs 411 mm in 2006, with 72% of this difference being attributed to April–May), potentially affecting the seasonal development of marsh vegetation. There were higher reflectance values for reed and club-rush beds in December 2005 than December 2004, while aquatic beds presented similar reflectance (Davranche, 2008). Moreover many reeds were still green in December 2005, while they are generally dry at that period. The values of NRMSE of significant parameters affecting model accuracy in reed beds recalled the correlation between stem diameter and leave size mentioned above. Percentage cover of apparent soil suggests that model power is lowest for extremely dense beds of high helophytes covering over 80% of dry ground.

170

A. Davranche et al. / Remote Sensing of Environment 138 (2013) 165–171

Very low transparency was linked to the worst model accuracies in aquatic beds. Hence, the model seemed to be less accurate for very turbid (b 25 cm) open marshes where reflectance values were increased in the green and red bands. Results in club-rush habitats were link to plant species. S. littoralis is much higher than B. maritimus. Its lower size and cover provides a patchy habitat favorable to the development of filamentous algae (classed in submerged species) near the water surface. Percent of apparent soil as an explaining factor is linked to one site where B. maritimus was mixed with Characea. Calcium carbonate contained in the Characea plants forms a white layer at the end of the senescence phase (due to phenology or dry up of the marshes) that increases reflectance values in the green and red bands. Differences in water and soil temperatures could also potentially explain the discriminatory power of the mid-IR band located closer to the thermal waveband spectrum. 4.3. Comparison of model accuracy, robustness and usefulness To our knowledge, no method has ever been developed to monitor water levels in wetlands at the pixel scale with such statistical prediction and robustness through space and time. Monitoring of water management using space-based data has actually focused on the detection of surface water level changes (Munyaneza, Wali, Uhlenbrook, Maskey, & Mlotha, 2009; Tan, Bi, Hu, & Liu, 2004; Wdowinski et al., 2008) or flooded areas (Oberstadler, Honsch, & Huth, 1997; Thorley, Clandillon, & De Fraipont, 1997; Töyrä et al., 2002). While the combination of SPOT and RADARSAT data remains more accurate for the detection of water presence through the vegetation (Töyrä et al., 2002), this study demonstrates that multispectral data can be powerful for fine scale monitoring of water levels through wetland vegetation. 5. Conclusion While the international community debates on the necessity to modify our understanding, behavior and practices relative to the management of water resources in a global warming context; our results offer innovant supporting tools to quantify such management and practices. Combined with previous results obtained for vegetation distribution (Davranche et al., 2010) and its state assessment (Poulin et al., 2010), it confirms the usefulness of multispectral data for long-term and cost-efficient monitoring of wetland resources. The high spatiotemporal variability in environmental conditions of Camargue marshes suggests this method could be useful for other wetland types as well, a hypothesis that is currently being tested in northern Europe. Our approach is also likely to perform well with other satellite data sources, should they provide a mid-infrared band. References Adell, C., & Puech, C. (2003). L'analyse spatiale des plans d'eau extraits par télédétection satellitale permet-elle de retrouver la marque cynégétique en Camargue? Bulletin de la Société Française de Photogrammétrie et de Télédétection, 172, 76–86. Aduah, M., Maathuis, B., & Hussin, Y. A. (2007). Synergistic use of optical and radar remote sensing for mapping and monitoring flooding system in Kafue flats wetland of southern Zambia. ISPRS commission, conference on information extraction from SAR and optical data with emphasis on developing countries, Istanbul 2007. Al-Khudhairy, D. H. A., Leemhuis, C., Hoffmann, V., Shepherd, I. M., Calaon, R., Thompson, J. R., et al. (2002). Monitoring wetland ditch water levels using Landsat TM and ground-based measurements. Photogrammetric Engineering and Remote Sensing, 68, 809–818. Aznar, J. -C., Dervieux, A., & Grillas, P. (2003). Association between aquatic vegetation and landscape indicators of human pressure. Wetlands, 23, 149–160. Bao, Y., Liu, L., & Wang, J. (2008). Estimating biophysical and biochemical parameters and yield of winter wheat based on LANDSAT TM images. IEEE Geoscience and Remote Sensing Symposium (pp. 7). Baret, F., Guyot, G., Begue, A., Maur, P., & Podaire, A. (1988). Complementarity of middleinfrared with visible and near-infrared reflectance for monitoring wheat canopies. Remote Sensing of Environment, 26, 213–225. Breiman, L., Friedman, J. H., Olshen, R., & Stone, C. (1984). Classification and regression trees. New York, USA: Chapman & Hall (358 pp.).

Caillaud, L., Guillaumont, B., & Manaud, F. (1991). Essai de discrimination des modes d'utilisation des marais maritimes par analyse multitemporelle d'images SPOT. Application aux marais maritimes du Centre Ouest. IFREMER report (H4.21), 485, (24 pp.). Chauvelon, P., Tournoud, M. G., & Sandoz, A. (2003). Integrated hydrological modelling of a managed coastal Mediterranean wetland (Rhone delta, France): Initial calibration. Hydrology and Earth System Sciences, 7, 123–131. Clay, D. E., Kim, K. -I., Chang, J., Clay, S. A., & Dalsted, K. (2006). Characterizing water and nitrogen stress in corn using remote sensing. Agronomy Journal, 98, 579–587. Davranche, A. (2008). Suivi de la gestion des zones humides camarguaises pat télédétection en référence à leur intérêt avifaunistique. (PhD Thesis): University of Provence (257 pp.). Davranche, A., Lefebvre, G., & Poulin, B. (2009). Radiometric normalization of multi-temporal SPOT 5 images for wetland monitoring: Accuracy of pseudo-invariant features vs. 6S atmospheric model. Photogrammetric Engineering and Remote Sensing, 75, 723–728. Davranche, A., Poulin, B., & Lefebvre, G. (2010). Wetland monitoring using classification trees and SPOT-5 seasonal time series. Remote Sensing of Environment, 114, 552–562. Delmeire, S. (1997). Use of ERS-1 data for the extraction of flooded areas. Hydrological Processes, 11, 1393–1396. Dervieux, A. (2005). La difficile gestion globale de l'eau en Camargue (France): le Contrat de delta. Vertigo, 3, (http://vertigo.revues.org/2411). Esposito, F., Malerba, D., Semeraro, G., & Tamma, V. (1999). The effects of pruning methods on the predictive accuracy of induced decision trees. Applied Stochastic Models in Business and Industry, 15, 277–299. Gao, B. G. (1996). NDWI—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment, 58, 257–266. Gond, V., Bartholome, E., Ouattara, F., Nonguierma, A., & Bado, L. (2004). Surveillance et cartographie des plans d'eau et des zones humides et inondables en régions arides avec l'instrument VEGETATION embarqué sur SPOT-4. International Journal of Remote Sensing, 25, 987–1004. Hanqiu, X. (2006). Modification of normalized difference water index (NDWI) to enhance open water features in remotely sensed imagery. International Journal of Remote Sensing, 27, 3025–3033. Hess, L. L., Melack, J. M., Novo, E. M., Barbosa, C., & Gastil, M. (2003). Dual-season mapping of wetland inundation and vegetation for the central Amazon basin. Remote Sensing of Environment, 87, 404–428. Holm, S. (1979). A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics, 6, 65–70. Horritt, M. S., & Mason, D. C. (2001). Flood boundary delineation from Synthetic Aperture Radar imagery using a statistical active contour model. International Journal of Remote Sensing, 22, 2489–2507. Huete, A.R. (1988). A soil-adjusted vegetation index (SAVI). Remote Sensing of Environment, 25, 295–309. Hunt, E. R., & Rock, B. N. (1989). Detection of changes in leaf water content using nearand middle-infrared reflectances. Remote Sensing of Environment, 30, 43–54. Jensen, J. R., Narumalani, S., Weatherbee, O., & Mackay, H. E. (1993). Measurement of seasonal and yearly cattail and waterlily changes using multidate SPOT panchromatic data. Photogrammetric Engineering and Remote Sensing, 59, 519–525. Lillesand, T. M., & Kiefer, R. W. (1987). Remote sensing and image interpretation (2nd ed.) New York: John Wiley and Sons (721 pp.). Mathevet, R., Le Page, C., Etienne, M., Lefebvre, G., Poulin, B., Gigot, G., et al. (2007). BUTORSTAR: A role-playing game for collective awareness of wise reedbed use. Simulation & Gaming, 38, 233–262. McFeeters, S. K. (1996). The use of the normalised difference water index (NDWI) in the delineation of open water features. International Journal of Remote Sensing, 17, 1425–1432. Munyaneza, O., Wali, U. G., Uhlenbrook, S., Maskey, S., & Mlotha, M. J. (2009). Water level monitoring using radar remote sensing data: Application to Lake Kivu, central Africa. Physics and Chemistry of the Earth, 34, 722–728. Oberstadler, R., Honsch, H., & Huth, D. (1997). Assessment of the mapping capabilities of ERS-1 SAR data for flood mapping: A case study in Germany. Hydrological Processes, 11, 1415–1425. Pearson, R. L., & Miller, L. D. (1972). Remote mapping of standing crop biomass for estimation of the productivity of the short-grass Prairie, Pawnee National Grasslands, Colorado. Processing of the 8th International Symposium on Remote Sensing of Environment (pp. 1357–1381). Ann Arbor, MI: ERIM. Pettigrew, W. T. (2004). Physiological consequences of moisture deficit in cotton. Crop Science, 44, 1265–1272. Poulin, B., Davranche, A., & Lefebvre, G. (2010). Ecological assessment of Phragmites australis wetlands using multi-season SPOT-5 scenes. Remote Sensing of Environment, 114, 1325–1628. Poulin, B., Lefebvre, G., & Mathevet, R. (2005). Habitat selection by booming bitterns Botaurus stellaris in French Mediterranean reed-beds. Oryx, 39, 265–274. Poulin, B., Lefebvre, G., & Mauchamp, A. (2002). Habitat requirements of passerines and reedbed management in southern France. Biological Conservation, 107, 315–325. Richardson, A. J., & Everitt, J. H. (1992). Using spectra vegetation indices to estimate rangeland productivity. Geocarto International, 1, 63–69. Rondeaux, G., Steven, M., & Baret, F. (1996). Optimization of soil-adjusted vegetation indices. Remote Sensing of Environment, 55, 95–107. Rouse, J. W., Haas, R. H., Schell, J. A., & Deering, D. W. (1973). Monitoring vegetation systems in the Great Plains with ERTS. Third ERTS Symposium, NASA SP-351, 1, . Sandoz, A., Chauvelon, P., Pichaud, M., & Buckwell, P. (2003). Estimation et suivi des superficies en eau par télédétection satellitale radar: résultats obtenus dans le delta du Rhône (France). Bulletin de la Société Française de Photogrammétrie et de Télédétection, 172, 69–75. Smith, L. C. (1997). Satellite remote sensing of river inundation area, stage, and discharge: A review. Hydrological Processes, 11, 1427–1439.

A. Davranche et al. / Remote Sensing of Environment 138 (2013) 165–171 Sokal, R. R., & Rohlf, F. J. (1995). Biometry: The principles and practices of statistics in biological research. NY: W.H. Freeman. Souza, R. P., Machado, E. C., Silav, J. A.B., Lagoa, A.M. M., & Silveira, J. A. G. (2004). Photosynthetic gas exchange, chlorophyll fluorescence, and some associated metabolic changes in cowpea (Vigna unguiculata) during water stress and recovery. Environmental and Experimental Botany, 51, 45–56. Spot Image (2008). Preprocessing levels and location accuracy. Technical information, www.spotimage.com Taft, O. W., Haig, S. M., & Kiilsgaard, C. (2004). Use of radar remote sensing (RADARSAT) to map winter wetland habitat for shorebirds in an agricultural landscape. Environmental Management, 33, 749–762. Tamisier, A., & Grillas, P. (1994). A review of habitat changes in the Camargue: An assessment of the effects of the loss of biological diversity on the wintering waterfowl community. Biological Conservation, 70, 39–47. Tan, Q., Bi, S., Hu, J., & Liu, Z. (2004). Measuring lake water level using multi-source remote sensing images combined with hydrological statistical data. Geoscience and Remote Sensing Symposium, 7. (pp. 4885–4888).

171

Therneau, T. M., & Atkinson, E. J. (1997). An introduction to recursive partitioning using the RPART routines. : Mayo Foundation (52 pp.). Thorley, N., Clandillon, S., & De Fraipont, P. (1997). The contribution of spaceborne SAR and optical data in monitoring flood events: Examples in northern and southern France. Hydrological Processes, 11, 1409–1413. Töyrä, J., Pietroniro, A., Martz, L. W., & Prowse, T. D. (2002). A multi-sensor approach to wetland flood monitoring. Hydrological Processes, 16, 1569–1581. Vermote, E., Tanre, D., Deuze, J. L., Herman, M., & Morcrette, J. J. (1997). Second simulation of the satellite signal in the Solar Spectrum, 6S: An overview. IEEE Transactions on Geoscience and Remote Sensing, 35, 675–686. Vescovo, L., & Gianelle, D. (2008). Using the MIR bands in vegetation indices for the estimation of grasslands biophysical parameters from satellite remote sensing in the Alps region of Trentino (Italy). Advances in Space Research, 41, 1764–1772. Wdowinski, S., Kim, S. W., Amelung, F., Dixon, T. H., Miralles-Wilhelm, F., & Sonenshein, R. (2008). Space-based detection of wetlands' surface water level changes from L-band SAR interferometry. Remote Sensing of Environment, 112, 681–696.