1. Introduction
Remote sensing provides valuable information about geographic spaces, allowing the detection of both naturally-occurring and anthropogenically-generated changes over wider areas than those commonly studied using data provided by field work. Different bands in satellite imagery sensing systems, with diverse spectral ranges, support highly differentiated applications.
Fluvial remote sensing is an emerging discipline joining river science remote sensing and geospatial sciences due to its high potential of reconstructing fluvial dynamics at a variety of time and spatial scales, and thanks to the increasing availability of remotely-sensed environmental data [
1,
2,
3]. Characterization of river hydromorphology greatly benefits from the exploitation of spatially- and temporally-explicit information provided by remote sensing techniques. Geospatial sciences, including geomatics and geostatistics, provide a wide set of tools that permit detection of water systems, such as rivers, and allow implementing models and processing procedures for the analysis of such resources. A wide range of river science topics and management applications has benefited from the use of satellite data, including river restoration [
4,
5,
6,
7], aquatic habitats characterization [
8,
9,
10,
11,
12], understanding biodiversity and ecosystem functioning of riverine environments [
12,
13,
14,
15], hazard mapping and river management at the catchment scale [
16,
17,
18], as well as mapping of morphological changes [
19,
20,
21]. In addition, river studies can take advantage of the use of Digital Elevation Models (DEMs) and aerial orthophotos [
11,
17,
22], where a more detailed dataset generally corresponds to a reduction of the covered area.
The availability of rich archives of satellite imagery is very useful in studying the time evolution of rivers’ hydromorphology, in particular for scarcely-anthropized systems with a high degree of naturality and in developing or emerging countries where there are often limitations in the availability of historical hydromorphological information. However, when a multi-temporal analysis is considered, it is necessary to take into account that different satellite missions acquire data with widely different spatial and spectral resolutions. Moreover, different light conditions have an impact on the ease of fusion of different imagery datasets even when a single mission is considered. Hence, the identification of an applicable procedure for the extraction of geometrical features of rivers, such as the active channel, is not straightforward. In fact, many studies are based on manual digitalization of geometrical features of interest [
23,
24]. Manual operations are however particularly costly, especially for broad study areas to be analyzed over time. Classification techniques are also applied to multi-band imagery to detect water and vegetation classes [
11,
25,
26]. In [
27], the active channel was detected by means of proprietary software implementing an object-oriented classification algorithm. As a result, the lack of an accurate, automatic, open source procedure able to detect and classify river features represents a limit in our present possibility to (i) exploit the variety of remotely-sensed data acquired by different sources and (ii) to merge them and reconstruct the evolutionary trajectory of river reaches.
In this work, a semi-automatic image processing procedure is presented to study river hydromorphological evolution on a timescale of several decades. The procedure involves the recognition of river reaches and the extrapolation of morpho-dynamic parameters including: the length and sinuosity index of main river centerlines, wet and active channel widths, surface coverage of water, vegetation and bare sediment units. At first, approximate river centerlines were extracted by processing imagery or DEM data. Active and wet channels were detected by means of a supervised classification of the above-mentioned surface coverage classes. Centerlines were extracted by means of Voronoi diagrams and further processed along with other quantities derived from transects of the surface coverage classified maps. The procedure was applied to a set of data including: DEM, high resolution orthophotos and satellite data from the CORONA, LANDSAT and Sentinel-2 missions. Satellite data cover the time period 1968–2017. The procedure was entirely implemented using free and open source software. Image processing was performed using GRASS (Geographic Resources Analysis Support System). GRASS is a free and open source Geographic Information System (GIS) for geospatial data management and analysis, image processing, graphics’ and maps’ production, spatial modeling and visualization. It is a founding member of the Open Source Geospatial Foundation (OSGeo). Other computations were made by means of ad hoc codes developed in the open source Python programming language.
The study concentrated on three highly dynamic rivers in Albania: the Shkumbin, Seman and Vjosë rivers. The rivers of interest present characteristic traits and marked natural evolution because of the limited anthropic pressures that have occurred over the past decades. Albania represents an area of high interest to investigate rivers’ hydromorphology under near pristine conditions (e.g., the Vjosë river) and their connections with natural (climate) and anthropic stressors, like deforestation, sediment mining in river channels and river regulation from hydropower, which have seen a marked increase in recent years [
28]. In addition, Albania represents a paradigmatic case for many emerging contexts where hydromorphological data availability is often limited and where the methodology presented in this work has therefore a high potential for enhancing quantitative knowledge of river forms and processes.
2. Study Site
Though published studies of Albanian rivers date back more than 40 years [
29,
30,
31,
32], relatively little information is available on their recent hydromorphological dynamics. Existing studies document the exceptional sediment transport rates of these streams [
33]. A comparison of selected Albanian rivers and some other European rivers (
Table 1) reveals that most Albanian rivers have much higher sediment transport rates per unit catchment area compared to other European rivers in similar climatic settings. This feature is particularly interesting in terms of river hydromorphology, as sediment transport rates represent one of the main control factors driving high rates of morphological activity in rivers [
34].
The identification of the river reaches of interest was performed by means of a combined analysis of the entire river course involving both aerial orthophotos and DEM. At first, the DEM was used to extract the river networks and hence to distinguish unconfined, partially-confined and confined reaches. Then, unconfined reaches were selected in order for the analysis to focus on reaches with the highest potential planform dynamics. Among all the identified reaches of potential interest, those presenting the greatest morphological differences between scenes acquired in 1968 and 2007 were selected by means of visual inspection. In the case of the Seman River, a scene centered on a large meander was taken. This meander reached its maximum curvature before a neck cut-off occurred within the study period. For the Vjosë River, a multi-thread section presenting high morphological dynamics was chosen. In the case of the Shkumbin River, a sufficiently extended scene of a single-thread reach in its lower course was attempted.
3. Data
To cover the longest possible time period, several imagery datasets were considered, including: Black and White (BW) images from the 1968 CORONA mission, RGB (Red, Green, Blue) aerial orthophotos from the 2007 Albanian National survey, multi-band LANDSAT imagery from 1972–2017 and 2017 multi-band Sentinel-2 imagery. Details concerning the imagery dataset will follow.
BW satellite images, 1968: digital scanned copies of photogrammetric BW films acquired in 1968 from one of the first generation of U.S. photo intelligence satellite systems, namely CORONA. Images were originally used by U.S. intelligence agencies and eventually declassified by Executive Order 12,951 in 1995 because of their potential usefulness in global change research. Images used in this study were in part downloaded for free and in part purchased from USGS (United States Geological Survey). The images were acquired in stereo mode by means of a system of two twin cameras named KeyHole (KH) 4A. The spatial resolution was about 2.8 m.
Mission data: 1049-2 12/22/1968; film IDs: DS1049-2155DA038, DS1049-2155DA042, DS1049-2155DF044, DS1049-2155DF045, DS1049-2155DA045.
RGB orthophoto, 2007: color orthophotos acquired in 2007. Images were geocoded in WGS84 (World Geodetic System 1984), and coordinates were given in a UTM cartographic plane. Images used in this study were saved from the Albanian State Authority for Geospatial Information (ASIG) portal using WMTS (Web Map Tile Services).
Multi-band LANDSAT imagery, 1972–2017: a set of multi-band LANDSAT imagery from 1972–2017. Considering the typical time scale of geomorphological phenomena involving natural river systems of interest, a scene every 5–7 years was considered. When possible, and in order to guarantee an acceptable presence of vegetation, scenes within the May–July period were selected. All LANDSAT data were downloaded for free from the Earth-Explorer website.
Table 2 contains details concerning LANDSAT data, including mission codes, sensing instruments, acquisition dates and spectral bands used in this work.
Multi-band Sentinel-2 imagery, 2017: Three tiles of the Level-1C product were considered. Each tile was acquired from the satellite S2A on 11 July 2017, at 9.30 UTC. All data of the COPERNICUS program were downloaded for free from the COPERNICUS HUB website. Bands used in the work had a spatial resolution of 10 m.
DEM: European DEM provided by EEA (European Environment Agency). The model was a hybrid product based on SRTM and ASTER GDEM data fused by a weighted averaging approach. The spatial resolution was 25 m.
The use of CORONA images was reported by [
35,
36,
37]. However, only recently, some authors have started to work with geometrically-corrected images [
38], even if a complete and rigorous procedure to ortho-rectify images was not yet available. In this work, the CORONA images served to establish reference information about river morphological features in order to compare the results of the temporal analysis. To enforce some geometrical improvements on the original uncorrected CORONA images, an ad hoc image rectification was performed estimating the parameters of a local transformation (second order polynomial) with a nearest neighborhood-based resampling. Coordinates of selected Ground Control Points (GCPs) were obtained from the aerial orthophoto. From the hydromorphological point of view, the reaches of the Shkumbin River considered in this study did not present significant differences from the characteristics of the reaches of the Seman and Vjosë rivers. Hence, the analysis based on the CORONA images was carried out only on the reaches of the Seman and Vjosë rivers. In
Table 3, the number of GCPs used and statistics about the Root Mean Square Errors (RMSEs) of the transformation are reported.
4. Analysis of Multi-Temporal Imagery Dataset
The adopted procedure was based on the following key steps: (a) identification of an approximate wet channel; (b) identification of a computational region built from the approximate river channel; (c) detection of active and wet channels along with vegetation and gravel coverage units on the active channel; (d) extraction of channel centerlines; (e) extraction of transects of the active and wet channels.
To tackle the differences in the spatial and spectral resolutions of the imagery dataset considered in this work properly, different strategies were developed for the implementation of the previous steps. A schematic description of the choices made for every kind of data is provided at the end of this section. In general, the requirements of user interaction were kept as low as possible to identify the most automated strategies possible. Whenever possible, key steps of the procedure were solved automatically by means of suitable techniques and tools, namely: color space transformation; pan sharpening; thresholding of vegetation and water indices; contextual supervised image classification of false color compositions, RGB and HIS bands; Voronoi diagrams.
Color space transformation: The RGB to HSL (Hue, Saturation, Lightness) color transformation was applied to facilitate the detection of the gravel class (see
Figure 1). The lightness component of the HSL bands corresponds to the arithmetic mean of the largest and smallest RGB components.
Pan sharpening: In scenes with LANDSAT 7 or LANDSAT 8 data, i.e., where the panchromatic band is present, pansharpening was performed.
Thresholding of vegetation and water indices. Plants have a typical absorption/reflection pattern; in fact, chlorophyll strongly absorbs visible light (0.4–0.7
m) for use in photosynthesis, whereas the cell structure of the leaves strongly reflects near-infrared light (from 0.7–1.1
m). The Normalized Difference Vegetation Index (NDVI) is a specific radiometric index that exploits such a typical absorption/reflection pattern, and it is used to map green leaf vegetation. Typically, the NDVI is computed according to the following equation:
where
and
R are the near infra-red and the red bands. NDVI values can range between −1 and +1. However, the typical range of NDVI values is between about −0.1 for non-vegetated surfaces and 0.9 for dense green vegetation canopies. The NDVI increases with increasing green biomass, changes seasonally and responds to favorable, e.g., abundant precipitation, or unfavorable climatic conditions, e.g., drought [
39].
Likewise NDVI, the Normalized Difference Water Index (NDWI) [
40] is a radiometric index that exploits a typical absorption/reflection pattern for water mapping purposes. After [
41], the NDWI is computed according to the following equation:
where
G and
are the green and the near infra-red bands.
Thresholds on the NDVI and NDWI maps were set in order to select only pixels corresponding to water. NDVI and NDWI histograms were studied, and pixels with value simultaneously greater than the 90th percentile of the NDWI map and lower than the 10th percentile of the NDVI map were labeled as water pixels. Once the initial water map was obtained, a buffer at a suitable distance was applied to ensure the inclusion of all parts of the wet channel in further processing. Afterwards, the enlarged water map was refined, excluding outside parts of the river reach, such as small pools or secondary channels outside the active channel by means of area thresholding. Likely, the vegetation map was produced selecting pixels with a value greater than the 80th percentile of the NDVI map; the gravel map was produced selecting pixels with a value simultaneously greater than the 95th percentile of the lightness map and lower than the fifth percentile of the saturation map. For every coverage class, the actual percentile values were slightly modified according to the specific imagery dataset considered.
Figure 2 and
Figure 3 show examples of NDVI and NDWI maps computed from Sentinel-2 imagery.
Contextual supervised image classification: To distinguish surface coverages of different types in multi-spectral imagery, a contextual classifier based on a Sequential Maximum A Posteriori (SMAP) estimation was used. The classifier models a spectral class by means of simple spectral mean and covariance as parameters of a Gaussian mixture distribution. The SMAP segmentation algorithm attempts to improve segmentation accuracy by segmenting the image into regions rather than segmenting each pixel separately. The SMAP algorithm exploits the fact that nearby pixels in an image are likely to have the same class. It works by segmenting the image at various scales or resolutions and using the coarse-scale segmentations to guide the finer scale segmentations. False-color composite maps facilitated the classification of water, vegetation and gravel classes. In
Figure 4, an example of the classified coverage units is shown; the orthophoto is used as a background image, and image contrast is reduced in order to enhance readability.
Voronoi diagrams: The channel centerlines were extracted starting from the Voronoi diagrams of the areas representing the active and wet channels. The centerline is represented by the polyline connecting the centers of all the highest circles inscribed in the Voronoi polygons. In
Figure 5, an example of the detected active channel and centerline is shown; the orthophoto is used as a background image, and image contrast is reduced in order to enhance readability.
From the active channel centerline, lastly, transects were created at a fixed distance of 200 m and with adequate lateral extension. From the channels’ centerlines and the transect sets, hydromorphological parameters were derived by means of an automatic procedure described in
Section 5.2.
Specific implementations done to adapt the procedure to different types of data are given hereafter in a step-by-step schematic way.
BW satellite images: Because of the very limited radiometric content and the reference role of BW images, the procedure, up to the extraction of channel centerlines step, was implemented entirely by hand.
RGB orthophoto and multi-band LANDSAT 1 and 2 imagery: DEM analysis for the extraction of river axis approximation; buffering approximate axis to limit the area of analysis; RGB to HSI color transformation for thresholding based initial gravel detection; buffering of detected gravel areas for the extraction of active channel; supervised contextual classification of active channel surface coverage; buffering and thresholding for coverage classes refinement; active channel identification after boundary vegetation removal.
Multi-band LANDSAT 3–8 and Sentinel-2 imagery: NDVI and NDWI analysis for the extraction of river axis approximation; buffering approximate axis to limit the area of analysis; RGB to HSI color transformation for thresholding-based initial gravel detection; buffering of detected gravel areas for the extraction of the active channel; supervised contextual classification of active channel surface coverage; buffering and thresholding for coverage classes refinement; active channel identification after boundary vegetation removal.
6. Discussion and Future Prospects
We discuss the outcomes of the present work at two levels: first at a technical, methodological level and then in relation to its broader significance for fluvial processes and related environmental implications for the examined case studies in Albania.
The proposed procedure proved to be a valid and effective tool to map river morphology and its temporal change. In particular, the extraction of centerlines, as well as active and wet channel width was sufficiently accurate, with differences with respect to the manual measurements below 6%. Moreover, the comparison with high resolution data (Sentinel-2) confirmed the accuracy of the results obtained by the Landsat imagery. Overall, the validation did not show any particular problem in the semi-automatic processing.
As satellite data were acquired by different sensors over a 40-year period, threshold values for radiometric indices were set specifically in order to account for different light conditions, the seasonality of the scenes and differences in spatial and spectral resolutions. The procedure provided high quality results for scenes obtained from the images acquired in the period 1986–2017, i.e., those with a spatial resolution of 30 and 15 m. As for the 1972–1981 data, the estimation of widths was affected by the relatively low spatial accuracy (about 60 m) of the scenes. However, the results obtained from the processing of CORONA images did not differ too much from those obtained from the processing of historical data. In the first years considered in this study, scenes were few in number and often in bad condition. In these cases, the homogeneity of the data concerning seasonality is lacking.
The general quality of the results and, in a particular way, the notable agreement between the outcomes derived from Sentinel-2 and LANDSAT 8 data prompt further applications of the procedure to entire river systems. Moreover, the high acquisition frequency of recent satellite missions offers the opportunity to implement rolling monitoring systems.
The proposed procedure exploits open data and is implemented by means of free and open source software. Such conditions can have great relevance for the application of the proposed methodology to those contexts where the availability of historical and recent environmental data is limited or of fragmented quality. This is the case, in general, of emerging or developing countries where rapid environmental changes driven by socio-economic development are presently occurring and expected in the near future.
In the specific case of the three examined Albanian rivers, one of the main outcomes of the analysis is that all the studied river reaches reduced their width, although at a different rate. The Vjosë River seems to be the least affected by this narrowing process, possibly thanks to its greater dynamism over time. The Seman and Shkumbin rivers experienced a remarkable channel narrowing, considering both the wet and even more the active width. They reduced their active width (mean value) to half or even one third of the value observed in 1972. This behavior is similar to what was observed for many other European rivers [
42], and in particular for many Italian rivers [
43]. Reduction of the active channel width can have dramatic consequences for a variety of aquatic, riparian and terrestrial ecosystems, which have direct or indirect legacies in relation to the morphological functioning of the river corridor [
44]. For instance, when the active river corridor width is reduced, the surface available for bare sediments and instream habitats is largely replaced by permanently-vegetated surfaces, which may impact entire bird communities that rely on bare sediment surfaces for nesting and incubation [
45] and fish communities that need to adapt to heavily-modified hydromorphological conditions [
46]. When habitat reduction exceeds a threshold, adaptation to the new river condition may no longer be possible, with potentially dramatic consequences in terms of biodiversity loss [
47]. Narrowing of the active river corridor may also negatively affect the hydraulic safety of human communities with respect flooding risk, because a narrowed river corridor determines higher water levels during floods. River narrowing is also often accompanied by riverbed incision [
48], which has often resulted in much increased damage for river structures like bridges, levees and check dams.
River width variability has often been observed at relatively slow time scales (decades [
48]), and it is therefore very important to develop tools that allow tracking such changes over the relevant time scales, and across a variety of geographical contexts characterized by different paths of data source availability. A complementary ingredient for effective environmental management would be to understand the causes and controls of those evolutionary river trajectories. Broadly speaking, the decrease of active channel width has been explained by the concurrent effect of different factors, including climate change, land use change, dam construction and in-channel gravel mining [
49]. The present study does not attempt to account for the causality of the observed changes on the selected Albanian rivers; for this, further research is needed. Here, we demonstrate that the developed procedure allows the collection of relevant data to reconstruct such changes in the river morphological trajectories. The ability to automatically detect the river planform at the catchment (or regional) scale is a first essential step towards effective management of such sensitive and dynamic landscape elements, within a broader perspective of environmental sustainability.
Future developments of the present work are both at a methodological level and in relation to its high potential for enhanced information acquisition on the temporal fluvial dynamics. From a technical point of view, the procedure could be improved or expanded by applying super resolution algorithms that could improve in a particular way the analysis of low resolution data [
50]. The CORONA images could be ortho-rectified to improve the metric information obtained from the processing of those data, thus greatly expanding the temporal window on which close multi-temporal comparisons are feasible. At the level of fluvial process understanding, of particular interest are applications of the proposed methodology to emerging or developing contexts where large river exploitation is planned at a global level [
51]. Integration with the analysis of the main anthropic and climatic controlling factors of river morphological evolution can provide a high added value for sustainable river management.