Introduction

River-floodplain systems contain complex habitat structures and are important for river management (Tockner et al. 2008). Habitat structures, described by land cover, habitat type, water qualities, and so forth, influence aquatic faunal communities (García et al. 2011). This is because the heterogeneity of habitat structures such as inflow, parapotamal, plesiopotamal, and palaeopotamal channels affects connectivity, which in turn enriches the diversity of aquatic species in a habitat (Tockner et al. 2010). Habitat structure also contributes to the overall species diversity, or scarcity, of aquatic fauna (Cornell and Lawton 1992; Tilman 1999). For example, the alteration of habitat structure has reduced the abundance of the Japanese eel Anguilla japonica Temminck and Schlegel, 1847 in rivers and lakes in Japan (Itakura et al. 2015). The turnover of habitats caused by disturbance, mainly flooding in rivers, is also important for aquatic diversity. Even though aquatic floodplain habitats are highly dynamic, small changes in the water level and flooding can lead to major habitat changes (van der Nat et al. 2003). Flooding increases similarity in species diversity among aquatic habitats in the river floodplain (Thomaz et al. 2007). However, flooding has resulted in minimal changes in the braiding, sinuosity, and aquatic habitat composition along the Fiume Tagliamento River, Italy (Arscott et al. 2002). Considering the highly dynamic nature of habitat diversity across the river-flood plain, proper management of habitat structure as well as ecological research are crucial for enriching species abundance and diversity (Osugi et al. 2007). Hinojosa-Huerta et al. (2013) showed that a reduction of surface flow led to a reduction in native tree cover but an increase in shrub cover in the riparian corridor delta of the Colorado River, Mexico without significant changes (< 30%) in overall vegetation. In addition, resilience of the dynamic nature of the riparian zone helps to restore its native form.

Researchers still largely rely on field surveys to describe and understand the spatial and temporal dynamics of habitat structures, despite constraints with respect to physical requirements, time, climate conditions, geographical limitations, and costs. Accordingly, there is little information on the dynamics of habitat structures at large spatial and temporal scales (Arscott et al. 2002; van der Nat et al. 2003) and high resolutions. Satellite images and remote-sensing techniques are increasingly used to monitor habitat structures owing to their wide spatiotemporal coverage, effectiveness for data collection, and repeatability (Mertes 2002; Henshaw et al. 2013; Moretto et al. 2013). For instance, high-resolution (0.5–5 m) satellite images have been used to estimate river depth (Legleiter et al. 2004; Kanno et al. 2011) and medium-resolution (10–15 m) satellite images have been used to estimate other water-quality parameters, such as the suspended sediment concentration (Jensen 1996; Xue et al. 2011).

Rosgen (1994) reviewed the classification system for natural rivers based on entrenchment, gradient, width/depth ratio, and sinuosity in various landforms. Satellite images are an important source for land cover classification and have been widely used for the characterization of fluvial habitat, for example, channel morphology and riparian land cover in the River Tay, Scotland (Bryant and Gilvear 1999), distribution and composition of aquatic insect communities effected by forest cover changes in Amazonian forests (Nessimian et al. 2008), aquatic and terrestrial ecosystems in the tundra wetland of the Lena River Delta, Northern Siberia (Schneider et al. 2009), and wetland in the Amazon Basin (Melack and Hess 2010). Changes in flow regime do affect fluvial and riparian habitats, as shown by alternation of habitat features in seasonal streams, and lesser physical alternation in the Sagura Basin, Spain in stable streams where there were no seasonal changes (Belmar et al. 2013). However, the major obstacle to remote-sensing techniques is the resolution, i.e., mainly the presence of shadows, clarity of water, variable substrates and target–sun angle geometry (Marcus and Fonstad 2008). There has been continual recent advancement in remote-sensing techniques as a tool for the classification of riverine habitats (Marcus and Fonstad 2010; Dufour et al. 2013; Whited et al. 2013; Romero et al. 2016).

Most of the proposed methods estimate a single habitat structure in rivers and thus have not been applied to a variety of physical habitats in river-floodplain systems. Henshaw et al. (2013) reported the successful development and application of multispectral Landsat TM satellite data comprising 15- to 30-m spatial resolution satellite images to monitor vegetation dynamics in a river floodplain, but little is known about the dynamics of habitat structures, especially habitat type classifications, at a large spatiotemporal scale and fine spatial resolution. The development of methods to accurately predict ecological responses, changes in biodiversity, and habitat degradation based on high-resolution satellite images would contribute to the design of effective mitigation measures.

Despite numerous studies on the dynamics of habitat structures in river floodplains, habitat type dynamics have not yet been examined at a large spatiotemporal scale. Habitat types have been monitored by field surveys at the reach scale (van der Nat et al. 2003), but more comprehensive monitoring of the spatial heterogeneity of habitat structures is needed. Furthermore, one issue with conventional methods for image classification is the need to select training data to obtain the pixel-value distribution. This selection needs to be done empirically and qualitatively to obtain sufficiently accurate estimates, but is sometimes tricky, as this procedure is based on trial and error. Hence, in the present study, a new method was developed to avoid issues associated with training data selection by converting pixel values into a digital number (DN) distribution. The main objective of this research was to develop a method for monitoring habitat structures in a river-floodplain system based on high-resolution satellite images and to apply this method to the floodplain in the middle reaches of the Tagliamento River, Italy. As a case study, the proposed method quantifies major land cover and aquatic habitat types in a floodplain to understand floodplain dynamics.

Materials and methods

Study area

The target area was the river-floodplain system of the Tagliamento River, located in Northeast Italy. The length of its main stream is approximately 170 km and it drains an area of approximately 2580 km2 from the Southern Alps to the Adriatic Sea (Tockner et al. 2003). It forms highly complex and dynamic river-floodplain systems on alluvial fans, and thus is considered the last morphologically intact river in Europe (Müller 1996; Ward et al. 1999; Tockner et al. 2003; Huber and Huggenberger 2015), though anthropogenic alterations have also been described in terms of hydrology and water quality in some sections (Tockner et al. 2008; Bertoldi et al. 2010). The land cover of the target floodplain consists mainly of five types: active riverbed, pasture, grassland, wetland, and cultivated field (Spaliviero 2003); however, the general land cover based on the map of the nineteenth century used in the classification of Spaliviero (2003) needs to be upgraded with the robust and latest information from high-resolution satellite images. Hence, for this study, land cover was categorized into three types and five habitats: river water, vegetated island and bare alluvium; channel, backwaters, ponds, vegetated island and bare alluvium (discussed in detail in “Categorization of habitat types”).

The Tagliamento River is characterized by an alpine to Mediterranean climate (Gurnell et al. 2001). November is the wettest and February the driest month (Gurnell et al. 2001). The mean annual precipitation is 1500–1800 mm in the northern catchment area and 1900–2900 mm in the alpine foreland area (Gunawardhana and Kazama 2012). Winter begins with snowfall, which starts in late November or early December, and the snow melts around late March. The river discharge is higher from April to June, and gradually decreases in the summer, probably with the minimum discharge in September. September is also the month with the highest amount of sunshine, with mean monthly sunshine of more than 200 h.

The unconstrained floodplain segments are characterized by a dynamic mosaic of aquatic/terrestrial habitats (Tockner et al. 1999; Arscott et al. 2002), a large number of vegetated islands (Edwards et al. 1999; Gurnell et al. 2001), and a high biodiversity of aquatic communities (Arscott et al. 2005). The floodplains show substantial morphological changes in the network configuration due to flood pulses and unconstrained river banks (Bertoldi et al. 2009). Interestingly, each flooding event along the Tagliamento River spatially reconfigures the habitat composition, but these events do not have substantial effects on the aerial proportion (i.e., composition) of each type of aquatic habitat in each floodplain, leaving a certain combination of major floodplain habitat types available for recolonization (Arscott et al. 2002; van der Nat et al. 2003).

To understand the continuous dynamics of habitat structures at a large spatial scale, the alluvial fan segment about 30 km from Venzone through Cornino and Villuza to Dignano was selected and was divided into five sections, A–E, from north to south (Fig. 1). In addition, three target reaches were selected within this segment: Cornino reach, Flagogna reach, and the reach downstream of Villuzza, referred to as reaches I, II, and III, respectively (Fig. 1). Reach I was about 1 km long and was located in section C, and II and III were about 1.5 km long and were located in sections D and E, respectively. Under base-flow conditions, reach I was dominated by exposed alluvium, with multiple channels (van der Nat et al. 2002). Reach II had a predominant wandering morphology with main channels and dead sub-channels. In the main channel, there were many longitudinal and lateral bars and islands, similar to reach I (Moretto et al. 2013). In reach III, surface water infiltrates the highly permeable alluvial aquifer, and a portion of this segment lacks surface flow (maximum dry length: 23 km) in low-flow conditions (Doering et al. 2007; Henshaw et al. 2013).

Fig. 1
figure 1

The target floodplain: the middle reaches of the Tagliamento River in Northeast Italy

Full size image

Satellite images

Cloud cover is an important parameter when selecting satellite images for better quality data. On manual sorting of various satellite images, lower cloud cover (< 1%) was found for the September. In addition, our target was to select images of the same month for different years having at least a 10-day antecedent dry weather period to minimize the impact of precipitation on changes in habitat structure. Hence, three orthocorrected RapidEye satellite images with a 5-m spatial resolution were obtained on 12 September 2010, 30 September 2011, and 15 September 2012. All images were obtained between 10:59 and 11:03 a.m., when the cloud cover was less than 1%. A dry weather period around 10 days prior to each date was confirmed, enabling us to assume that the dynamic equilibrium of the river network was not affected by a probable flood due to precipitation. All 16-bit RapidEye satellite images consisted of four spectral bands, namely, blue, green, red, and near-infrared (NIR), with wavelengths of 440–510, 520–590, 690–730, and 760–920 nm, respectively.

Land cover classification

The maximum-likelihood classification (MLC; hereafter referred to as MLC-1), one of the most widely used methods for land cover classification, was applied to the supervised image classification (Wernick and Morris 1988); the ISODATA method (ISO) was applied to the unsupervised classification (Ball and Hall 1965; Smith et al. 2008). Several input images are required for land cover classification for both the ISO and MLC. In ISO, the pixel-value distribution of each land cover type is estimated by automatic clustering, while MLC requires training data, which are used to obtain the pixel-value distribution for each class. Generally, the accuracy is lower for ISO than for MLC using the same input images, but MLC requires reliable training data for acceptable accuracy (Smith et al. 2008).

To avoid the difficulty associated with the selection of training data in supervised classifications, in this study, a new method for quantitative image classification based on ISO was developed. In this method, when multi-spectral band images are used as input data, DN distributions for several bands can be clustered into classes because each land cover type is characterized by a different DN distribution (Bruzzone and Prieto 2001). In the conventional ISO method, hereafter referred to as ISO-1, overlapping plots and linear relationships among bands often make it difficult to accurately identify land cover classes. The proposed method for image classification, hereafter referred as ISO-2, was designed to avoid such problems in clustering by pixel-value transformation (Kojima 2011). The major improvement of this method is that the DN distribution for each class is separated owing to the increased accuracy of the ISO estimates, which minimizes overlap in the DN distribution of each class between bands in the transformed domain.

Suppose that n classes of land cover are classified based on a satellite image with M bands. The DN distribution of M bands for class i is expressed using the function fi as follows:

$$ f_{i} \left( {{\text{DN}}_{1} ,{\text{DN}}_{2} , \ldots {\text{DN}}_{J} , \ldots ,{\text{DN}}_{M} } \right) + \emptyset_{i,i}^{'} = 0, $$
(1)

where \( {\text{DN}}_{M} \) is the DN of class i in band M and \( \emptyset_{i,i}^{'} \) is the error term equal to the deviance. According to this relation, the DN distribution of class i in band J is expressed as follows:

$$ {\text{DN}}_{J} \left( i \right) = g_{i} \left( {{\text{DN}}_{1} ,{\text{DN}}_{2} , \ldots ,{\text{DN}}_{M} } \right) + \emptyset_{i,i}^{'} {\text{for}} M \ne J, $$
(2)

When Eq. (2) is applied to all pixels in an image (i.e., other land cover classes), the error term differs for each class and can be expressed as a matrix as follows:

$$ \emptyset '_{i} = \left[ {\begin{array}{*{20}c} {\begin{array}{*{20}c} {\emptyset '_{i,1} \cdots {\text{on Landcover}} \left( 1 \right)} \\ \vdots \\ \end{array} } \\ {\emptyset^{\prime}_{i,i} \cdots {\text{on Landcover }}\left( i \right)} \\ \vdots \\ {\emptyset^{\prime}_{i,n} \cdots {\text{on Landcover}} \left( n \right)} \\ \end{array} } \right], $$
(3)

where \( \emptyset '_{i} \) is the deviance distribution consisting of the errors of each class \( \emptyset^{\prime}_{i,n} \) obtained by applying Eq. (2) to class n.

If the image is reproduced with the deviance \( \emptyset '_{i} \) to the pixel value, instead of DN, the deviance attributed to class i, i.e., only \( \emptyset '_{i,i} \), will be closer to zero than that of other classes, since \( \emptyset '_{i} \) is the error from the estimation equation of the DN distribution of class i. Finally, for all classes from 1 to n, the error number and DN are calculated to acquire the pixel-value distribution for each class between two images for image classification.

Application of the new classification method

To validate the DN distribution in four bands for each class, the true-color images were reproduced by combining red, green, and blue band images, and DN was selected from 30 pixels for each land cover type from this true-color image. Referring to the RapidEye sample data obtained in 2012, the DN distribution for each class was plotted and fitted to a linear spectral unmixing model for green and other bands, except the NIR band on river water (Fig. 2). Similar trends were found for each class and the other RapidEye images could be represented as follows:

$$ {\text{DN}}_{\text{grn}} \left( i \right) = \left( {a * {\text{DN}}_{\text{blue}} \left( i \right) + b*{\text{DN}}_{\text{red}} \left( i \right) + c*{\text{DN}}_{\text{NIR}} \left( i \right) + d} \right) + \emptyset_{i}^{'} , $$
(4)
Fig. 2
figure 2

The relationships of digital number distributions between green and blue, red, and near-infrared (NIR) for each land cover class of 2012 RapidEye satellite images: a river water, b vegetation, and c bare alluvium. The samples were taken from 50–100 randomly selected points of each land cover type indicated in the floodplain map (d). The figure parts indicate that the relationships among bands can significantly differ according to land cover type

Full size image

where \( {\text{DN}}_{\text{grn}} \left( i \right) \), \( {\text{DN}}_{\text{blue}} \left( i \right) \), \( {\text{DN}}_{\text{red}} \left( i \right) \), and \( {\text{DN}}_{\text{NIR}} \left( i \right) \) are the observed DN distributions of class i in green, blue, red, and NIR bands, respectively, and a, b, c, and d are constants. The error was estimated by applying Eq. (4) to the sample data for river water, vegetation, and bare alluvium and was used in the following ISO image classification.

In this study, image classification was conducted using GRASSGIS (version 6.4.2) for all three classification methods, namely, the conventional MLC (MLC-1), the conventional ISO (ISO-1), and the new ISO method developed in this study (ISO-2).

To check the accuracy, the omission rate, commission rate, and κ-coefficients were calculated by cross-validation using reach II. For this validation, around 50% of all pixels for each land cover were taken from true-color images. Each coefficient was estimated as follows:

$$ {\text{Omission rate}} = \frac{1}{{n_{i} }}Q_{ij} \left( {i \ne j} \right), $$
(5)
$$ {\text{Commission rate}} = \frac{1}{{n_{i} }}Q_{ji} \left( {i \ne j} \right), $$
(6)
$$ {\text{Kappa coefficient}} = \frac{{N\mathop \sum \nolimits_{i = 1}^{m} Q_{ii} - \mathop \sum \nolimits_{i = 1}^{m} \left( {Q_{i + } \times Q_{ + i} } \right)}}{{N^{2} \mathop \sum \nolimits_{i = 1}^{m} \left( {Q_{i + } \times Q_{ + i} } \right)}}, $$
(7)

where \( n_{i} \) is the number of pixels in class i for the validation data, \( Q_{ij} \) is the number of pixels belonging to class I but wrongly assigned to class j in the validation data, \( Q_{ii} \) is those pixels that are correctly classified, \( Q_{i + } \) is the total pixels of class i in the validation data, \( Q_{ + i} \) is the total of class i in the classification result, \( N \) is the number of pixels in the whole target image, and m is the number of classes. If κ is greater than 0.80, the accuracy of the estimate is deemed sufficient (Cohen 1960; Kojima 2011).

Categorization of habitat types

The habitats in the Tagliamento River have been categorized into seven types: channel, alluvial channel, backwater, pond, bare alluvium, riparian forest, and vegetated island (van der Nat et al. 2003). As it was not possible to clearly distinguish between alluvial channels, gravel bed and bare alluvium from the RapidEye satellite images at 5-m spatial resolution, all these three habitat types were categorized as bare alluvium. Hence, habitats in the river floodplain surrounded by riparian forest were classified into five types, namely, channels, backwaters, ponds, vegetation islands, and bare alluvium (Fig. 3).

Fig. 3
figure 3

Categorization of land cover and habitat types of the Tagliamento River

Full size image

For the fluvial habitats, ponds were defined as pools of water within the river floodplain that were not connected to any streams. Backwaters were connected to streams at only the downstream or upstream end, and the remaining river water land cover could be classified as channel. Based on the land cover estimation using satellite images, the river network was obtained; thereby, the fluvial habitat was classified morphologically into channels, backwaters, and ponds.

Estimation of the temporal shift in habitat structure

Referring to the habitat type categorization based on satellite images, the area of each habitat can be calculated as follows:

$$ A_{t1} \left( i \right) = R_{s}^{2} *N_{t1} \left( i \right), $$
(8)

where \( A_{t1} \left( i \right) \) is the area of habitat I in period t1 (squared meters),\( R_{s} \) is the resolution of satellite images (5 m in this study), and \( N_{t1} \left( i \right) \) is the number of pixels in the habitat of interest at time t1 on the satellite images.

To examine the temporal habitat changes after several flooding events, the conversion rate to the next time step t2 was represented as follows:

$$ C_{t1,t2} \left( {i,j} \right) = A_{t2} \left( {i,j} \right)/A_{t1} \left( i \right), $$
(9)

where, \( C_{t1,t2} \left( {i,j} \right) \) is the conversion rate of habitat type i to j between the period from t1 to t2, \( A_{t2} \left( {i,j} \right) \) is the area of habitat type j converted from type i, and \( A_{t1} \left( i \right) \) is the area of habitat type i at time t1. In this study, the conversion rate was calculated for each habitat change from 2010 to 2011 and from 2011 to 2012 in three reaches.

To understand the habitat changes (or lack thereof) over several years, the rate of unchanged habitat for each type was also calculated by overlaying the habitat type classification images for each year using the following function:

$$ U_{\text{tn}} \left( i \right) = A_{t1} \left( i \right)\mathop \cap \nolimits A_{t2} \left( i \right)\mathop \cap \nolimits \cdots \mathop \cap \nolimits A_{tn} \left( i \right) /A_{t1} \left( i \right), $$
(10)

where \( U_{tn} \left( i \right) \) represents unchanged habitat type i over the period from t1 to tn.

Results and discussion

Land cover classification

All parameters of Eq. (4) were obtained from the DN distribution estimates for four bands as shown in Table 1 and Fig. 2. The correlations between each estimate and observation exceeded 0.90. High-resolution satellite images showed higher correlation coefficients and the highest value was obtained for bare alluvium, with a maximum of 0.99 (Fig. 2). The DN distribution of 50–100 randomly selected pixels of Fig. 2d is reflected in Fig. 2a–c, which was obtained after merging the frequency distribution of river water, vegetation and bare alluvium in the blue, red, green and NIR bands (Supplementary Fig. S1), and the frequency distribution of the blue, red, green and NIR bands for river water, vegetation and bare alluvium for the 2012 image (Supplementary Fig. S2).

Table 1 Parameter estimates for the digital number correlation equation
Full size table

According to the pixel-value distributions for the RapidEye satellite image obtained in 2012, the distributions of each class partially overlapped using both ISO-1 and ISO-2 (Fig. 4). However, for ISO-1, the DN distributions of river water, vegetation, and bare alluvium totally overlapped, as DNs were plotted on the same line for the blue, green, and red bands, whereas ISO-2 did not show such an overlap. In addition, the pixel-value distributions for ISO-2 were closer to zero than those of ISO-1 when the estimation class was objective and the distributions were the same (Fig. 4). The plots of each land cover type for ISO-1 were similar, especially among neighboring bands, except for red and NIR. This means that the classes cannot be effectively separated and thus linear plots make clustering difficult.

Fig. 4
figure 4

Pixel-value distribution in input images for 2012 using a the conventional ISODATA (ISO-1) and b the proposed ISODATA method (ISO-2) (the same numbers of pixel values were selected for comparison between a and b)

Full size image

The differences in image classification based on the pixel-value distribution among input images, overlapping for ISO-1 and less overlapping (or no overlapping) for ISO-2, were also supported by the higher κ-coefficient for ISO-2 (Table 2). The κ-coefficient was greater than 0.89, exceeding a reference threshold of 0.80, indicating better efficiency for ISO-2 for all images obtained in 2010, 2011, and 2012. Hence, in this study, the distribution of pixels based on the newly developed method showed a better ability to isolate river water, vegetation, and bare alluvium land covers.

Table 2 Commission rate, omission rate, and κ-coefficient of land cover classification for each year
Full size table

In the land cover classification results, the κ-coefficient exceeded 0.83, a sufficient threshold value, for ISO-1, ISO-2, and MLC-1 (Table 2). The κ-values for ISO-2 were 0.89, 0.98, and 0.95 for 2010, 2011, and 2012, respectively. Similarly, the κ-values for MLC-1 were 0.87, 0.97, and 0.95 for 2010, 2011, and 2012, respectively. The values of κ from the ISO-1 classification were 0.83, 0.95, and 0.93 for the same years, respectively.

The outputs obtained by image classification for ISO-1 indicated that the river water data were overestimated and the boundaries between each land cover type were ambiguous (Fig. 4). The boundaries among each land cover were clear in ISO-2 and MLC-1 (Fig. 4). The results for vegetation differed between ISO-2 and MLC-1. ISO-2 was apparently able to estimate vegetation at a finer scale than MLC-1, especially in 2010 and 2011 (Fig. 4). Similar trends to those in reach II were found in reach I and III.

Therefore, the developed method (ISO-2) enables more accurate estimates of the land cover class without a repetitive data-selection training process than the conventional unsupervised method (ISO-1), as the pixel values were clearly differentiable among river water, vegetation, and bare alluvium in the Tagliamento River. In addition, the accuracy of the newly developed method was high, i.e., greater than 89%. However, it is necessary to validate the efficiency of this new method for other land cover classes (e.g., urban areas), and the efficiency for the classification of types of soil, vegetation, and plants should be evaluated in further studies.

Spatiotemporal land cover changes

Referring to the land cover classification results, the land cover ratio and the unchanged area in the floodplain surrounded by riparian forests were obtained from 2010 to 2012 in the middle reaches of the Tagliamento River (Fig. 5). The presence of vegetation was relatively low in areas A and B in the upper part of the Tagliamento River and higher in area D, the middle part, which includes reach II. The presence of river water was higher in the middle region (namely, areas C, D, and E, which include reach I and II) than in other regions (Fig. 5a). The rate of unchanged river water in each area was also calculated (Fig. 5b) and was higher in area D for all periods from 2010 to 2011, 2011 to 2012, and 2010 to 2012, probably due to the presence of dead subchannels in reach II (Moretto et al. 2013). In all areas, the unchanged area was greater for 2011 to 2012 than 2010 to 2011, and in the lower part (areas D, E, and F), the unchanged area decreased in the downstream direction for all periods.

Fig. 5
figure 5

Outputs obtained by the image classification of ISO-1, ISO-2, and conventional maximum-likelihood classification (MLC-1) from 2010 to 2012 in reach II, and true-color RapidImage in 2012 (top right)

Full size image

Based on the composition of habitat types obtained from the land cover classification, the dominant habitat type was bare alluvium, covering 55–75% of each reach (Fig. 6). The unchanged area was highest for vegetated islands (19.4–26.5%) in reach II, and lowest (7.5–8.7%) in reach III. Table 3 shows the annual habitat turnover from 2010 to 2011 and 2011 to 2012 in reaches I, II, and III, respectively. The unchanged area of fluvial habitats, such as channels, backwaters, and ponds, exhibited random patterns in the three reaches. For instance, the unchanged area of channels was highest in reach I (31.7 and 55.7%), II (44.6 and 64.5%), and III (64 and 56.8%) for 2010 to 2011 and 2011 to 2012 (Table 3).

Fig. 6a, b
figure 6

Temporal shift of each habitat. a Composition of each land cover type in the middle of the Tagliamento River, and b unchanged river water in the middle of the Tagliamento River

Full size image
Table 3 Annual habitat turnover from 2010 to 2011 and from 2011 to 2012
Full size table

The observed turnover was explained by two main factors, i.e., flooding, which usually occurs around ten times per year in the Tagliamento River, and gradual changes associated with the seasonal changes in discharge. These are both important processes determining habitat properties and faunal species richness (Benke et al. 2000; Tockner et al. 2000). Since the RapidEye images used in this study were obtained within a half-month period each year, the extrapolation of the present results to other seasons should be done with care, as seasonal variation in vegetation has been reported in an earlier study in the Tagliamento River (Doering et al. 2007).

The unchanged area was highest for bare alluvium in reaches I and III and for vegetated islands in reach II (Fig. 6). For fluvial habitats, especially in channels, the unchanged area was highest in reach II. This is explained by the high proportion of unchanged area in the subchannel in reach II, i.e., 75% from 2010 to 2011 and 66% over 2 years, compared with the averages of 33 and 20% for all reaches, respectively. These surviving patches may serve as refuges for benthic macroinvertebrates during floods or increasing discharge. Such refuges play important roles in the recolonization of newly formed habitats after a flood (Sedell et al. 1990; Robertson et al. 1995; Winterbottom et al. 1997; Rempel et al. 1999).

Variability in habitat abundance and channel complexity is associated with changes in water and sediment supply, flooding, and vegetation-driven stabilization processes (Kiss et al. 2008). According to the results of this study, the habitat distribution exhibits spatial and temporal changes, especially for fluvial habitats, while the composition of habitat types is roughly constant in the middle reaches of the Tagliamento. These findings are consistent with those of previous studies (Arscott et al. 2002; van der Nat et al. 2003). The high-resolution satellite image classification system developed in this study enables the continuous monitoring of the spatiotemporal dynamics of habitat structures at a larger spatial scale and a higher accuracy (> 89%) than those of field observations (Fig. 7).

Fig. 7
figure 7

Unchanged rate (%) of each habitat type from 2010 to 2012

Full size image

Conclusion

The dynamics of habitat structures in floodplains located in the middle reaches of the Tagliamento River were estimated using high-resolution satellite images, i.e., multi-spectral images of RapidEye at a spatial resolution of 5 m. This estimation facilitates discussion of the annual turnover of floodplain habitat structure from 2010 to 2012. A major advantage of the image classification method developed in this study is that it does not require empirical training with data and additional observations while covering the changes in habitat structure at such a large spatiotemporal scale with high resolution; hence, it is relatively simple, easy to implement, and accurate (> 89%) compared to the conventional method. The specific conclusions of this study are as follows.

  • Three classes of land cover in the Tagliamento River floodplain were identified: river water, vegetation, and bare alluvium.

  • The unchanged area of vegetation was relatively small in the lower part and bigger in the middle part of the river, and unchanged river water was higher in the middle part of the Tagliamento River. In all areas, the rate of unchanged habitat was greater for 2011–2012 than for 2010–2011.

  • The dominant habitat type was bare alluvium, covering 55–75% of each reach. The rate of unchanged vegetated islands was highest (19.4–26.5%) in reach II and lowest (7.5–8.7%) in reach III. Unchanged habitat was highest for bare alluvium and vegetated islands.

  • The habitat distribution exhibited spatial and temporal changes, especially for fluvial habitats. The composition of habitat type was preserved in the middle reaches of the Tagliamento River.

  • For future research, since the resolution and observation frequency of satellite images are constantly improving, the detailed and comprehensive monitoring of habitat structures, especially in urban areas, and the classification of soil, vegetation, and plant types are recommended. More effective river-floodplain monitoring could be developed by integrating the accurate method developed in this study, images with a higher resolution, and field observations.