1. Introduction
Ecosystem monitoring with replicable remotely-sensed methods offers the distinct advantage of repeated, homogeneous coverage of large areas, with little extra effort [
1,
2,
3,
4]. This allows the development of time series datasets at coherent spatial scale irrespective of site accessibility. Application of remotely-sensed techniques for wetland mapping and monitoring has received a lot of attention [
5,
6,
7] due to this ecosystem’s decline and contribution to human well-being [
8,
9]. Wetland classifications have been performed with a multitude of sensors (aerial, multispectral, and synthetic aperture radar SAR) under a wide array of parametric and non-parametric statistical approaches using pixel- and object-based algorithms [
5,
10,
11,
12,
13,
14]. Among spectral bands, the near infrared (NIR) and red edge (RE) have been identified as the most useful for delineating wetland types [
5,
6,
7,
11,
15,
16], along with short-wave infrared (SWIR) bands, which are sensitive to both soil and vegetation moisture [
6,
17]. Thermal infrared (TIR) bands have also been used successfully to distinguish water bodies from vegetation and soil covers [
6,
7], as well as for identifying inundated wetlands [
6,
18]. With microwave bands, optimal values for incidence angles, wavelengths, and polarizations differ according to wetland vegetation types, with longer wavelengths performing better in forested wetlands [
14]. What stands out from the abundant literature reviews on the remote mapping of wetlands is that, owing to the diversity of their vegetation morphologies, which are highly dynamic and often hard to discriminate from that of terrestrial ecosystems, there is no standard methodology to map wetlands on a large scale [
6,
19]. However, because hydroperiod is a prime factor influencing biodiversity and the services provided by aquatic ecosystems [
20,
21], surface water area is often used as a proxy in remote sensing to identify wetlands or estimate spatio-temporal changes in their extent [
22,
23,
24,
25].
Although supervised classifications based on spectral analysis have been useful for accurately and repeatedly mapping water bodies [
26,
27,
28], the application of spectral indices has gained popularity because they are considered less restrictive and more reproducible, especially for applications at large or on global scale [
4,
29]. Several spectral indices have been developed to monitor surface water areas using satellite imagery [
25,
30,
31,
32,
33,
34,
35,
36,
37]. They generally use the near infrared (NIR) and/or short-wave infrared (SWIR) bands because water absorbs most radiation at NIR wavelengths and beyond, in contrast to other landscape features [
33,
38]. Their increasing applications under various situations has led to several modifications to improve classification accuracy, especially relative to the misclassification of turbid waters [
39] or the noise caused by built-up land and shadow [
22,
40]. For instance, the NIR band in the Normalized Difference Water Index (NDWI) developed in 1996 [
37] was replaced by the SWIR band in 2006 to reduce disturbances related to built-up lands giving rise to the Modified NDWI [
36]. Under the same reasoning, it was further suggested to calculate the MNDWI using the band SWIR2 instead of SWIR1 [
22]. An Automated Water Extraction Index (AWEI) has been proposed under two versions to reduce misclassifications related to either shadow or built-up land [
31]. A comprehensive comparison of the performance of these water indices using Landsat scenes from Australia revealed that most indices tend to underestimate water presence, being affected by water color and the presence of non-water features in a pixel [
41]. Considering the unique spectral characteristics of water bodies in the visible and infrared wavelengths, the application of fixed thresholds to spectral bands remains a valuable approach for delineating aquatic ecosystems [
40,
42]. Arguments against threshold-based methods is that they do not necessarily perform as well outside the areas where they were developed [
25]. Although water indices are considered as more stable because they use band ratios, recent studies have revealed similar shortcomings when water maps are confronted with ground-truth data, imposing the use of specific thresholds (different from 0) to increase classification accuracy [
22,
41].
Because of its particular climate characterized by an annual water deficit, many wetlands of the Mediterranean basin are flooded only seasonally [
43]. In this area, wetlands colonized by reeds, bulrushes, and other emergent plants provide sheltered refuges for wildlife and primary resources for industry and local populations [
44]. The biodiversity and socio-economic value of these wetlands primarily rely on the timing and duration of inundation [
45]. Increased water stress predicted under climate change projections [
46,
47,
48,
49] will negatively affect ecosystem services (provision of food, building materials, recreational activities, etc.) and biodiversity (e.g., reduction of suitable feeding, spawning, nesting and nursery grounds to birds, amphibians and fish) [
8,
50,
51]. Accordingly, failure to detect water presence under vegetation could lead to errors in the (1) classification of wetland habitats; (2) detection of changes in wetland functions; (3) assessment in water resource use, availability or management; and (4) extrapolation of wetland biodiversity and services [
19,
46,
52].
Detection of water under wetland vegetation has received little attention in the development of remote sensing algorithms and indices using optical data [
25,
26]. Vegetation growth inhibits optical sensors in variable ways, depending on the plant species, by interfering with water detection [
25,
26,
53,
54]. With radar sensors, emergent vegetation presents differences in surface roughness and increases the amount of backscattered radiation from inundated surfaces, making the discrimination of land and vegetated wetlands problematic [
24,
54]. Long wavelength SAR sensors with small incidence angles can penetrate vegetation more successfully, but the signal that is partially blocked by the vegetation creates a specific backscatter response due to double-bounce scattering [
55]. Because vegetation growth and structure will induce different scattering mechanisms [
54,
55,
56,
57], the ability to detect surface water will vary across space and time being influenced by vegetation morphology and phenology [
14,
58].
Several studies have recently tested the performance of standard spectral indices for surface water detection under various situations in terms of terrain and sensor [
22,
40,
41,
59]. Capitalizing on a solid ground-truth sample, this study aims at identifying what are currently the best options for detecting surface water, with special attention to water under dense vegetation cover. However, instead of individually testing the water and non-water classification accuracy of each index relative to ground-truth data as previously done [
22,
40,
41,
59], this study uses a data mining approach to identify what performs best among water indices, vegetation indices, and spectral bands used alone or in combination, using decision trees as classifiers. This work was carried out with the optical sensors of Sentinel 2, Landsat 7, and Landsat 8. The recent launch of Sentinel 2 satellites provides scenes of relevant spectral, spatial, and temporal resolution for monitoring wetlands routinely and at no cost. Although Landsat data have lower temporal and spatial resolutions, they were also selected because of their exceptional data archive that enables long-term trend assessments.
4. Discussion
In contrast to current water indices, the logical rule presented in this paper for detecting Water In Wetlands (WIW) performs equally well in the absence or presence of vegetation above the water surface. Seasonal wetlands are ecologically and economically important ecosystems that are particularly sensitive to climate change [
46]. A robust tool for monitoring annual and seasonal trends in their hydrology is needed by practitioners interested in the conservation of these vulnerable ecosystems because hydrology is a prime factor affecting their biodiversity and contribution to humankind [
8,
43]
Although based on independent field and remotely-sensed data in differing proportions, as well as various satellite sensors and different time periods, dichotomous partitioning with Landsat 7, Landsat 8, and Sentinel 2 led to the same logical rule for predicting water presence. In all cases, the near-infrared band (NIR) was first selected, followed by the second shortwave-infrared band (SWIR2). None of the standard water indices found in the literature were selected by our models for detecting water in wetlands. Although an increased performance of water indices has been obtained recently by adding specific threshold values or by using them in combination with vegetation indices [
22,
41,
59], applying simple threshold values to the NIR and SWIR2 bands appeared to provide better results in this study. Switching the threshold values among satellites or using different atmospheric correction methods provided similar water maps, suggest that our approach is robust and replicable. Furthermore, when applied to a Sentinel 2 scene of the Doñana marshes in southern Spain [
79], the WIW logical rules provides a Kappa coefficient of similar value (0.84) to the one obtained for the Camargue wetlands. Overall, Sentinel 2 scenes systematically provide better classifications, presumably because of their higher spatial resolution compared to Landsat sensors. All models performed better than the water index previously developed in Camargue with Spot 5 which used a combination of green and SWIR wavelengths (MIFW index, [
26]). The Spot-5 sensor had a single SWIR band that was located in the lower wavelength (1.58–1.75 μm), corresponding to the SWIR1 of the satellites used in this study and not selected in our classifiers.
A closer look at the classification tree reveals that areas which reflect heat back into the atmosphere such as dry ground with little (e.g., mud flats) or no (e.g., road, buildings) vegetation are discarded by the reflectance values of the near infrared radiation. In a second step, the combined action of the short infrared penetrating the vegetation, its absorption by water and reflection by the ground is useful for identifying flooded areas, even under vegetation cover. Penetration of the NIR and SWIR wavelengths (800 and 2700 nm) through organic matter has several applications [
80]. It has been shown that SWIR penetration capacity increases with increasing wavelengths [
81,
82]. In our case, it seems that the SWIR band behaves as in reflectography, the process used to highlight charcoal drawing underlying master paintings. According to this technique, a light source is used to illuminate the painting and the SWIR passes through the paint, being reflected by the canvas and absorbed by the charcoal. The optimal wavelength for passing through all paint layers is around 2 μm [
83], similarly to the SWIR2 bands selected in the WIW logical rule. It is noteworthy to mention that the only existing water index that can detect water under vegetation (MNDWI2) also uses SWIR2 wavelengths.
The main types of flooded vegetation in the Camargue correspond to grasses (e.g., rice), succulent shrubs (e.g.,
Arthrocnemum,
Salicornia,
Salsola), trees (
Tamarix sp.), and beds of emergent plants having variable height and density such as
Ludwigia spp., club-rush, rush, sedge, fen-sedge, and common reed [
84]. Based on visual interpretation of the water maps, the WIW logical rule performs equally well with all these types of vegetation. Apart from the MNDWI2 that can detect water in the early stage of vegetation growth, all water indices tested in this study failed to detect water under all types of vegetation cover. The particular case of hunting reed marshes is interesting because all water indices could detect permanent water in areas free of vegetation that are managed for ducks, but none of them could detect water into the reeds surrounding these pools which are also flooded for most of the year. Likewise, water under halophilous scrubs, which are a common habitat in Camargue, went completely undetected in all water indices.
Standard water indices applied to our dataset had nevertheless Kappa coefficients generally above 0.6. Kappa coefficients are considered as the most robust method to measure classification accuracy because they take into account the possibility of the agreement occurring by chance alone. However, their calculation remains limited to the reference and validation points provided by the observer. Accordingly, a high Kappa coefficient does not necessarily mean that the map is accurate. Originally, only optical-space derived data (from SPOT 5 to Landsat 8 and then other satellites, see methods) were used in this study to develop the WIW equation. Such an approach provided good overall accuracy and Kappa coefficients but the resulting water maps were wrong when confronted with ground-truth knowledge. Conventional statistical methods are not designed to deal with erroneous data. When using inaccurate training data, misclassifications (e.g., sampling points for which the reflectance value is located outside the confidence interval of the studied variable) are discarded from the original group of data. This contributes to reducing the confidence intervals of the original dataset and the data marginally correctly classified are suppressed to optimize good classifications. While such procedure gives a high potential for good statistical results, it amplifies the original model’s flaws. In our case, when the original model correctly classified water presence or absence only half the time for a specific land cover class, the following water model then systematically misclassified this land cover type. Our solution was to add ground-truth data on water presence in the training sample for those land cover classes that were identified as providing false results. This allowed us to restore the original confidence interval of the dataset by increasing the number of points lying outside its limit values. This approach provided satisfactory results because the water maps were coherent with reality and the final models provided high classification accuracy.
With all satellites, uneven slopes facing north of a small mountain range located outside the study area were misclassified as flooded during the winter months. Since all satellites were passing over the Camargue in the late morning (between 10:10 and 10:40 CET), this confusion is probably associated with shadows caused by the winter sun that is too low to light the northern face of the mountains. This problem has previously been reported with most water detection methods and can be solved by combining spectral indices or adding elevation data [
23,
31]. On a few occasions, some permanent waters in a large and deep lagoon were identified as dry by all satellites. Considering that these scenes were systematically acquired under the condition of strong winds (> 100 km/h), these artifacts were probably caused by strong waves causing foam on the water surface. Such phenomenon is, however, unlikely to occur in shallow or seasonal wetlands.
5. Conclusions
The Camargue or Rhône delta comprises a high diversity of natural and human-modified habitats. The method developed in this study for detecting Water In Wetland (WIW) is hence likely to be applicable to many other wetland areas, especially around the Mediterranean Basin where similar types of landscapes are found. It would be interesting to test its performance under subtropical and tropical climates where wetland vegetation is more luxuriant and stratified, such as in the Everglades [
25]. Automated methods for defining optimal thresholds would certainly increase performance of the WIW of which the main strength is to rely on high shortwave infrared wavelengths (SWIR2). Considering that our models were transferable from one satellite to the other, it seems likely that they would perform equally well with other satellites should they have SWIR and NIR sensors of comparable wavelengths. The model developed with Landsat 7 is probably the most robust for use with other satellites given its high performance with Landsat 8 and Sentinel 2, which is attributed to the wider acceptance range of its NIR sensor. Considering that Landsat 5 uses exactly the same sensors as Landsat 7, application of the WIW logical rule will permit territorial planners, wetland managers, and environmental scientists to follow water dynamics back to 35 years ago and, hopefully, for many years into the future with Sentinel 2, Landsat 8, and other satellites.
The definition of wetlands provided by the Ramsar Convention is very inclusive [
85]: “…wetlands are areas of marsh, fen, peatland or water, whether natural or artificial, permanent or temporary, with water that is static or flowing, fresh, brackish or salt, including areas of marine water the depth of which at low tide does not exceed six metres.” Considering the high temporal resolution of Sentinel 2 scenes (every 5 days), cumulative water maps built with the WIW logical rule could further be used for mapping a wide range of wetlands which are either periodically or permanently flooded. Such approach could be a good substitute to wetland mapping based on their vegetation characteristics and would further enable the monitoring of hydrology, in addition to wetland extent and location. Flooding dynamics have important implications for multiple services provided by wetlands (e.g., flood mitigation, water purification, wildlife habitat, and recreational potential), including carbon and methane cycling [
8,
86].