Basin‐wide hydromorphological analysis of ephemeral streams using machine learning algorithms ‡

Sustainable river management now encompasses a much wider concept that includes hydromorphological and fluvial habitat studies. In ephemeral streams, the geomorphological characterization of channels is complex due to episodic flows and riparian vegetation dynamics. Stream channel survey and classification at the watershed scale provide the basis for geomorphological conservation, process interpretation, assessing sensitivity to disturbance, and identifying reaches that supply and store sediment. Here, we present a stream classification based on a two‐step approach: (1) automatic river segmentation based on spatial variability in channel/valley morphology from topographic measurements (LiDAR, light, detection and ranging), and (2) fluvial landform and vegetation density mapping derived from multispectral open‐source satellite images (Sentinel‐2) using support vector machine (SVM) and Random Forest (RF) algorithms. These analyses provide continuous, quantitative spatial values of geometric (channel/valley width, slope gradient, and route distance), landform (active channel and gravel bars with five densities of vegetation cover), and hydraulic (specific stream power) variables. Four stream types were identified in the Rambla de la Viuda catchment (~1500 km2), an ephemeral gravel‐bed river in eastern Spain. The spatial distribution of channel types is explained by differences in geometry (active channel width, valley width, and slope gradient) and a hydraulic parameter (specific stream power). The landforms/vegetation patterns provided insight on causal relationships between erosion and deposition processes during high flow periods and the time since the most recent large disruptive flood event. Channel type distribution provided first‐order predictions about the location of reaches that supply and store sediment and thus information on sediment continuity along the river. Dam effects on downstream reaches resulted in geomorphological disequilibrium, producing narrowing of the active channel, slope reduction, and a decrease of gravel bar areal extension. The proposed catchment scale analysis provides a comprehensive and replicable methodology for environmental planning in Mediterranean ephemeral streams to guide further hydromorphological surveys at the reach scale.


| INTRODUCTION
Intermittent rivers and ephemeral streams (IRES; Datry et al., 2017) account for over 50% of the total length of the global river network (Skoulikidis et al., 2017) and comprise a large portion of Mediterranean watersheds. The extension of Mediterranean IRES is expected to increase in the near future, as some perennial rivers are experiencing intermittency due to climate change, land-use changes, and water demands (Gallart et al., 2012). Over the last few years, increasing efforts have been made to update methods for the operational monitoring, assessment, and classification of the water regime, biodiversity, and ecological functions of IRES (Datry et al., 2014;Datry et al., 2018;Leigh et al., 2016). These studies recognize the importance of the biophysical structure (e.g., channel morphology, substrate composition) and the eco-hydrological processes that support ecosystem functions (Datry et al., 2018). In dryland regions, physical and biological processes are more diverse and complex than their perennial counterparts due to inherent differences in climate, hydrology, geology, physiography, and human activities (Jaeger et al., 2017). Therefore, the assessment of their hydrological regime, network connectivity and geomorphological conditions is a challenging task (Gallart et al., 2017;Skoulikidis et al., 2017).
Recently, river management has adopted a hydromorphological approach [e.g., European Union (EU) Water Framework Directive, European Commission, 2000, 2003 to evaluate integrated aspects of geomorphology, hydrology, and freshwater ecology in preserving the important ecosystem services that rivers provide. Following a topdown watershed analysis, several frameworks for generating processbased understanding have been proposed for perennial rivers, namely the River Styles Framework (Brierley & Fryirs, 2005), IDRAIM (Rinaldi et al., 2015), or the multi-scale hierarchical framework  with all of them requiring a river segmentation frequently based on expert criteria before river characterization. However, specific approaches for developing process-based understanding of hydromorphological conditions of ephemeral streams are still scarce (e.g., IHG-E in CHJ, 2018) or mostly related with the evaluation of the biological component of the fluvial system (i.e., TREHS in Gallart et al., 2017). The hydromorphological approaches to stream segmentation are critical in assessing the environmental conditions of Mediterranean ephemeral streams and is the necessary stage for the consequent understanding of geomorphological dynamics across different spatial and temporal scales.
Ideally, a hydromorphological assessment should include a catchment morphological survey to evaluate the spatial and temporal links between flow regime and landform as part of the river's environmental characteristics (England & Gurnell, 2016). In most cases, existing geomorphological surveys are limited and catchment-scale data collection becomes costly (Downs & Thorne, 1996). As such, a Catchment Baseline Survey (CBS) is typically derived from the technique of geomorphological mapping with the support of topographic maps and aerial photographs (Cooke & Doornkamp, 1990;Hooke, 1997;Grabowski & Gurnell, 2016;Sear & Newson, 2010). Basin-wide analyses preferably include a combination of measurements of channel geometry, riparian land uses, channel dynamics, and evidence of previous management activity (Downs & Brookes, 1994). In perennial rivers, these surveys may be representative of channel conditions in the order of 10 to 100 years (Downs & Thorne, 1996), whereas in ephemeral streams with landforms controlled by episodic flooding the inferred interpretations may be limited to much shorter periods (Sanchis-Ibor et al., 2019). New approaches are therefore required to provide continuous, efficient, and replicable geomorphological data surveys at the catchment scale (Demarchi et al., 2020;Notebaert & Piégay, 2013;Piégay et al., 2020). Currently available open-access geospatial data, global positioning systems, and multispectral airborne and satellite images (e.g., Landsat, Sentinel) provide suitable datasets for repeated analysis at the catchment scale. For instance, a Pan-European riverscape unit mapping was recently produced based on Copernicus VHR Image Mosaic (2.5 m pixel) and the EU Digital Elevation Model (DEM) (25 m grid), although it is only applicable to large rivers (Demarchi et al., 2020). What is missing from this general riverscape analysis, however, is a hierarchical context for stream reach classification based on quantitative geometric and geomorphological variables at the catchment level. A need therefore exists for a methodological approach for the objective spatial analysis of Mediterranean ephemeral networks that could be used as a decision support tool to assess geomorphological functionality, response to disturbance, and connectivity along the river network.
The main objective of this article is to develop a basin-wide approach for ephemeral streams in support of sustainable river management, based on stream channel classification and open-source geospatial data. The specific objectives include: (1) to use objective criteria with a replicable methodology; (2) to provide objective delineation of functional river segments along the longitudinal continuum; (3) to generate an inventory of morphological and vegetation conditions at a reach scale that will subsequently be used to train machine learning algorithms; (4) to classify river channels based on their geometric, morphological and hydraulic characteristics; and (5) to identify river segments with degraded geomorphological conditions due to natural or human disturbance.

| STUDY AREA
Our study area is the Rambla de la Viuda, a Mediterranean ephemeral river that drains a catchment of $1500 km 2 in eastern Spain The climate is meso-Mediterranean in the upper catchment, which reaches 1250-1720 m above sea level (a.s.l.) (pine and evergreen oak forests), and thermo-Mediterranean (shrublands) in the rest of the catchment, with a mean annual temperature of 9 to 15 C, daily winter minima below 0 C and summer maxima above 30 C. Precipitation varies from north to south across the basin, from annual values of $750 mm/a in the Iberian Range to $400 mm/a in the lower catchment near the Mediterranean Sea (Mateu, 1974). The main rainfall season occurs in autumn (September-November), with a second maximum in late winter to spring (March-May).
The discharge pattern is strongly marked by the seasonal rainfall regime, the high bedrock permeability due to the dominance of karstified limestones, and the high transmission losses affected by hyporheic flow interactions within wide Quaternary alluvial fill basins (Beneyto et al., 2020). Thus streamflow occurs during 31 d/yr on average, typically lasting 2-3 days, and stream runoff requires an accumulated rainfall episode larger than 70 mm (Camarasa & Segura, 2001). Intense rainfalls (i.e., up to 300 mm in 24 h) occur mainly in the autumn and are associated with mesoscale convective cells (Llasat & Puigcerver, 1990), eventually producing large floods with peak discharges in excess of 1500 m 3 /s at Maria Cristina Reservoir Machado et al., 2017). These floods contribute to this irregular hydrological regime, producing up to 80% of the annual discharge volume (Segura & Camarasa, 1996).

| METHODOLOGY
The methodology involves seven main phases ( Figure 2): (1) preprocessing of satellite data from Sentinel-2 (i.e., calculation of spectral and textural indices); (2) manual collection of landform sample data for variable selection and supervised classification); (3) selection of spectral bands or indices for performing supervised classification; (4) data production from extraction of geomorphic variables using a geographic information system (GIS); (5) application of statistical methods to a set of selected independent variables to detect internally homogeneous segments significantly different from the adjacent segments; (6) cluster analysis of river segments based on geomorphic variables to identify river channel types; and (7) interpretation of fluvial geomorphology and stream functionality of the segment types (T) derived from the objective classification method.

| Data sources
3.1.1 | Remote sensing variables extraction from Sentinel-2 In order to develop our land cover analysis, a total of 16 remote sensing variables based on previous literature about land cover and land-use analysis (Borak, 1999;Marceau et al., 1990) were selected: six single spectral bands of Sentinel-2; three spectral indices; and seven textural indices. All of these data were cloud processed and downloaded from the Google Earth Engine (GEE) platform.
We performed the analysis using Sentinel-2 imagery with the advantage of its finer spatial resolution (10 m) in comparison to other open data satellite multispectral sensors (e.g., Landsat -30 m). For this study, we selected Red, Green, Blue (RGB), NIR1, SWIR1, and SWIR2 as candidate variables with the two latter being resampled to 10 m using the bilinear interpolation method. We used a Level 2A surface reflectance product, filtering by date (15 June 2019) and tile (T30TYK). We selected this image because it had less than 10% cloud cover.   , 1974) and Green Red Vegetation Index (GRVI) (Tucker, 1979).
Furthermore, the Normalized Difference Water Index (NDWI) (Gao, 1996) was calculated providing a measure of water content in soils and vegetation.
Beyond spectral indices, textural indices describe the distribution of grey tones in an image. This consists of a grey level co-occurrence matrix (GLCM) that interprets the spatial distribution of pairs separated by a certain distance in a given direction (Haralick et al., 1973).
Seven texture metrics were selected in our study as candidate variables: variance, correlation, contrast, entropy, second moment, mean, and dissimilarity (Haralick et al., 1973). These metrics have been successfully applied by previous authors for vegetation, soil, crop and building classification (Leinenkugel et al., 2019;Zhang & Zhu, 2011). 3.1.2 | Sample points for machine learning Classification of landforms in Mediterranean ephemeral rivers was carried out manually to construct a landform sample dataset for variable selection and supervised classification according to seven categories, and was based on the classification of Sanchis-Ibor et al. (2017), namely (1) channel and unvegetated gravel bar area (i.e., unvegetated gravel bars); (2) barely vegetated area, or deposits covered by grasses and scattered shrubs (< 5%); (3) mixed vegetated area covered by shrubs (< 50%) and scattered trees (< 5%); (4) fully vegetated area covered by shrubs (75%) and scattered trees (< 25%); (5) tree covered alluvial areas, covered by compact masses of trees in alluvial areas; (6) agricultural land; (7) water body, namely refers to reservoir and pond areas.

| Remote sensing variable selection
With a total of 16 variables calculated, the variable selection was carried out according to relative importance metrics in landform sample dataset. This step is required to obtain acceptable computational times and to reduce variable overlapping. We applied the Random Forest (RF) algorithm to rank variable importance in sample data using the Gini Impurity index (Breiman, 2001). RF is a machine learning algorithm based on tree decision rules (Breiman, 2001). The nodes are divided, in each decision tree, using a randomized subset of the variables. The final outcome is the majority votes from all trees. To obtain a consistent measure of the impurity of each variable we implement and additional k-fold cross-validation (k = 3) splitting the original sample into three (i.e., two training (66%) and one test (33%) datasets), repeated 10 times. This procedure generates 30 measurements of variable impurity. The aggregated median value of impurity was used to rank the variable importance. Additionally, RF requires hyperparameters optimization; for this reason, we performed a 10-fold repeated cross-validation and the number of predictors randomly selected in each tree decision (mtry) was 1-7, the minimum size of a node to be divided (min.node.size) was 2-30, and the number of trees value (ntree) was 500. This process was carried out with the CARET package (Kuhn, 2020)   3.2 | River elements classification modelling River elements classification was tested using two machine learning algorithms, support vector machine (SVM) and RF. RF algorithm was previously used to asses variable importance (Section 3.1.3.). In this section we use the results of the importance ranking of each variable to select the independent variables applied in the SVM and RF models. Radial SVM is a supervised machine learning technique that separates different categories by means of hyperplanes (Cortes & Vapnik, 1995).
Model training was the same for both algorithms. The sample point dataset was split into 80% for model training and 20% for testing model predictions. The best hyperparameters combination was tested using a more exhaustive cross-validation strategy, leaveone-out-cross-validation (LOOCV). This method is an iterative method that starts by using all available observations as a training set except one, which is excluded for validation. The process is repeated as many times as observations are available, excluding in each iteration a different observation, fitting the model with the rest and calculating the error with that observation. Finally, the error estimated by LOOCV is the average of all the calculated errors.
SVM also requires hyperparameter optimization. In SVM, we optimized the gamma (defines how far the influence of a single training example reaches) and cost (trades off correct classification of training examples against maximization of the decision function's margin) hyperparameters with values spaced logarithmically between 0.001-1 and 1-700, respectively. For the SVM prediction we used a radial basis function kernel due to its high performance in land cover classifications (Thanh Noi & Kappas, 2018). For RF, using a linear spacing, the mtry parameter ranged from 1 to 3, the min. node.size from 2 to 30, and the ntree value was 500. To optimize hyperparameters and to model in both approaches we used the CARET package (Kuhn, 2020).
The supervised classification of river elements was validated using the testing subset. Additionally, we calculated the confusion matrix showing the performance of pixels with correct and incorrect predictions of the SVM and RF models. An accuracy index was obtained from the ratio of the correct predictions to the total number of pixels in the test dataset. Furthermore, we used the Kappa Cohen index (Cohen, 1960) when evaluating the classifiers' prediction performance.

| Geomorphic variables
Five geomorphic variables were used to characterize the Rambla de la Viuda stream network: active channel width, valley bottom width, slope gradient, route distance, and specific stream power ( Table 1)

| Methods for river segmentation
The detection of homogeneous river segments based on statistical methods has been developed extensively in the last few decades In the present work, multivariate analyses by multi-response permutation procedures (MRPP, Mielke, 1991) were performed to segment the studied fluvial network. MRPP are non-parametric techniques that allow the river network to be classified into internally T A B L E 1 Geomorphic variables used for achieving river segmentation by multi-response permutation procedures.

| Channel type classification
After applying MRPP, each river segment was characterized by: (a)  where significant differences existed, the differences were assessed by pairs using a Mann-Whitney Wilcoxon test (Wilcoxon, 1945  Table S1) shows that agricultural land is the class that contained the most misclassified pixels for both models. Based on these quantitative comparisons of the models' predictions, the SVM model was selected for the landform analysis. SVM shows the best performance, although all classes were predicted well by both models (Figure 3b,c). Among the classes, the best classification was obtained for the tree covered alluvial area class. The prediction map (Figure 3) confirms that most misclassification occurred in the agricultural land class, which contained pixels misclassified as mixed vegetated or fully vegetated (Figures 3b and 4e). The channel and unvegetated gravel bar class was properly detected by the SVM method, which avoided any misclassification patterns or incorrectly classified pixels (Figure 4d,e).

| River segmentation using multivariate methods
A total of 62 segments were detected using the multivariate method ( Figure 4a, see Table S2 for a complete description of the 62 segments). The average segment length is 4 km, with length reaches ranging between 0.6 km and 19.2 km in length. The highest number of segments was detected for the Rambla de la Viuda River corridor (36.8 km in total length), where 14 segments were detected with an average length of 2.6 km (Figures 4a and 5a,b), whereas only two segments were differentiated for the Riu Alcora (10 km in total length) with an average length of 5 km. Other streams (dels Olles, Benafigos and Solana) were considered as individual segments because significant differences along the stream were not detected using the multivariate method (Figure 4a). will support that the segmentation process results in a robust and geomorphologically-meaningful segmentation method.

| Typological classification of segments
The analysis of the geomorphic parameters resulted in the differentiation of four main types of segments based on hierarchical cluster analysis (see Section 3.5., Figure S3). Results are shown as segment types T1 to T4 in Figure 4(a). T1 and T2 comprise the highest number of segments with 19 and 21, respectively, whereas T3 and T4 contain 12 and 10 segments, respectively (Table S3).
T1 is characterized by narrow active channels (14.7 m on average), significantly narrower than other groups (Willcoxon test, p < 0.05), within a narrow valley bottom (98.7 m on average), and with slope gradient and average specific stream power higher than other groups ( Figure 6; Table S3). Channel and unvegetated gravel bars, barely vegetated area, and mixed vegetated area landforms show a significantly lower extent compared to other groups, whereas the relative extent of the fully vegetated area and, in particular, tree covered alluvial area were significantly higher than other groups (Figure 7).
This group is mainly located in the headwaters in mountain areas and in bedrock confined valleys, which explains the frequent occurrence of relatively narrow channels occupied by mature vegetation (Figure 4a,b).
In T2, active channel width (32.9 m on average) is significantly wider than in T1 and significantly narrower than in T3 and T4 (Wilcoxon test, p < 0.05). Nevertheless, these channels are located within relatively narrow valleys (75.3 m width on average) ( Figure 6; and tree covered alluvial area are significantly lower than in T1 and T2. In T3, the segments contain a well delineated and continuous active channel partly laterally confined within mixed bedrock-alluvial valleys (Figures 4d and 7).
Finally, T4 is characterized by a significantly wider active channel and valley bottom (Wilcoxon test, p < 0.05) than other groups, namely of 217.3 m and 369.7 m width on average, respectively, leading to the lowest values for specific stream power ( Figure 6). In T4 the proportion of channel and unvegetated gravel bar extents are similar to in T2 and T3. However, T4 shows relatively wide and continuous channels, with a higher extent of mixed vegetated area and agricultural land due to their location on wide alluvial valley bottoms (Figures 4e and 7).

| DISCUSSION
Our methodological approach for ephemeral stream classification combines two independent quantitative geomorphologic analyses and provides two distinct advantages: (1) the use of objective criteria with a replicable methodology to define stream channel types, and (2)

| Landform class mapping challenges
Our approach, which combines remote sensing and machine learning techniques, provides objective and semi-automatic fluvial landform mapping for Mediterranean ephemeral streams. Recent advances in geomorphologic mapping and spatial classification of rivers are based on aerial images (Gilvear et al., 2004), high-resolution aerial images Rivas Casado et al., 2015, 2017, very high resolution (VHR) or VHR near-infrared aerial imagery combined with topography (Demarchi et al., 2016(Demarchi et al., , 2017(Demarchi et al., , 2020, and satellite images combining DEM and orthophotos (Spada et al., 2018 et al., 2020). However, using VHR images to this extent increases the computational time but also results in an excessive amount of information that requires a simplification process to portray fluvial units at the map scale. We propose the use of Sentinel-2 satellite images because of their 12 spectral bands and great potential for diachronical analysis covering specific dates, which is crucial for documenting flood disturbance and human pressures. Moreover, Sentinel-2 images require a relatively low computational load, which in this case was achieved using Intel Iris Plus Graphics, 16GB RAM and Intel Core i7-1065G7 CPU wherein the processing time is 30 min for ranking variable selection and 120 min for completing the machine learning algorithms. Despite its limited resolution (10 m), which hinders its application in narrow streams, the worldwide availability of Sentinel-2 images makes our approach extendable to basically any river and, in particular, to IRES.
Remote sensing has also benefited from technical developments in big data analysis. In our study, machine learning algorithms (SVM and RF) were applied to obtain a spatial classification. Using the same combination of single spectral bands and indices in both methods, we obtained accuracies and kappa values higher than 80%, with the best accuracies obtained with the SVM method as shown in previous studies (e.g., Phiri et al., 2020).

| Segmenting and establishing types for ephemeral channels
River segmentation from multivariate analysis combined with the classification of six landform units from remote sensing has resulted in the channel network being classified into four types. This channel network classification based on automatic river segmentation and remote sensing (RS) contains features of three previous geomorphic classifications that describe: (1) longitudinal stream zonation within drainage basins, namely erosion, transport and deposition zones (Bull, 1979;Schumm, 1977;Sutfin et al., 2014); (2) fluvial landform assemblages (Montgomery & Buffington, 1997;Rosgen, 1994); and (3) process domains (Montgomery, 1999;Whiting & Bradley, 1993). The spatial data analysis combines geomorphological and vegetation attributes, providing three main advantages. Firstly, the multivariate segmentation solves the difficulty of objectively identifying channel types along bedrock-alluvial transitional states through analysis of DEM data. This segmentation procedure identifies segment boundaries derived from changes in geometric and hydraulic variables that otherwise would not be straightforward from a visual or field interpretation, particularly in transitional reaches (Sutfin et al., 2014). Secondly, segment-level landform units were mapped using machine learning algorithms on satellite images, which offers uniformity and consistency in the mapping criteria. This technique enables very efficient application to large areas with moderate effort, even in those areas located in remote sites that are difficult to access, as long as a satellite image is available.
Thirdly, the statistical cluster analysis of geomorphologic assemblages at the segment scale provides a suite of process domain attributes for segment affinity among channel morphological types. Different statistical approaches for hierarchical clustering have been used in fluvial geomorphology, including the Ward's method applied in this study (Clubb et al., 2019;Henshaw et al., 2020), and multivariable k-means statistical clustering (Dallaire et al., 2019;Thoms et al., 2018). The kmeans cluster approach provides a simple set of clusters without a particular structure within them whereas Ward's method finds a large set of clusters which are merged in the process based on their affinity (Ward, 1963). Ward's method is the most widely used algorithm for identifying landscape units, and differs from other methods in that it uses an analysis of variance approach to determine distances between clusters. In general, this method is very efficient, as it allows the distances between clusters to be evaluated from an analysis of the variance (Zhang et al., 2017).
The variables involved in the cluster analysis include both RS derived landform units and topographic/hydraulic variables inherently connected by fluvial processes. Commonly, these process domains (Montgomery, 1999) are associated with landscape controls idealized as a downstream progression from source (headwater: bedrock-control), transport (piedmont: mixed bedrock-alluvium) and response reaches (low lands: alluvial streams), analogous to those described by Sutfin et al. (2014). In our case, the spatial variability of landforms and process domains are complex due to geologic and structural controls (i.e., horst and graben zones). Bedrock type and topography strongly influence the in-channel landform structure and variability (Grant & Swanson, 1995) in response to the dominant process and capacity for geomorphic adjustment (Fryirs & Brierley, 2010). This classification approach therefore reflects fundamental differences along the river channel network and recognizes disruptions of the river continuum of a morpho-sedimentary character along a gradation of stream power (Bridge, 1993) and confinement (Sutfin et al., 2014) trends.
The spatial distribution of channel types can be explained by differences in geometric (i.e., active channel/valley bottom width, and slope gradient) and hydraulic (i.e., specific stream power) parameters.
The alluvial landforms' (mainly classified by vegetation cover) distribution and extent depend on erosion and deposition spots during flow runs and on the time since the large disruptive flood event.
There is a general spatial trend from headwater to lowland in the distribution of channel types, namely from T1 to T4 (Figure 8). Headwater segments (i.e., T1 and T2) are characterized by confined valleys that allow direct sediment input via hillslope processes. Colluvial channels, within source areas dominated by gravity flows, were not identified at the satellite image scale. In T1 segments, the specific stream power is very high but the abundance of tree covered alluvial area is likely related to the episodic rate of sediment supply particle transport conditions (Papangelakis & Hassan, 2016). T2 segments show functional geomorphic characteristics compatible with a seasonal flow that efficiently conveys sediment load producing significant geomorphological adjustments (Montgomery & Buffington, 1997). This channel type is morphologically resilient to textural and geometric changes (Fryirs & Brierley, 2016) with increased sediment connectivity during peaks of gully and hillslopes material input.
T3 segments comprise mixed alluvial-bedrock channels whose spatial distribution is controlled by geologic and structural conditions  (Montgomery & Buffington, 1997;Schumm, 1977). The highest percentage of mixed vegetated areas is characteristic of multithread channels that shift, producing local erosion and deposition (Schumm, 1977). In general terms, segments of T4 are competence-limited channels (lowest specific stream power values, Figure 6), which implies only a partial connection to the stored sediment.

| Linking reach-level landform classes and disturbance regime
A challenge in watershed-based classifications is the spatial-temporal geomorphic adjustment of river landforms to either natural or anthropic geomorphic disturbances (Fryirs & Brierley, 2016). In ephemeral streams, most geomorphic effective work (i.e., sediment transport and landform changes) is performed by large flood events (Tooth, 2000). Indeed, the distribution of vegetation species across the landforms is governed by the tolerance and exposure of species to flood disturbance and water stress regimes (Hupp & Osterkamp, 1996;Manning et al., 2020). According to Montgomery (1999), a watershed may be divided into spatially identifiable areas, or process domains, along the river network, dominated by distinctive geomorphic processes, disturbance regimes, and response potential. The landform and vegetation structure of segment types provides insight into the processes resulting from the disturbance regime and their physical impacts (Figure 8). For instance, disturbance by floods is enhanced in confined channels in comparison to unconfined channels (Gregory et al., 1991). The landform distribution analysis confirms that the partially confined streams in T3 comprise the highest unvegetated and barely vegetated landform areas, likely associated with frequent vegetation reset by flooding (Hupp & Osterkamp, 1996;Sanchis-Ibor et al, 2017). The greatest percentage of mixed vegetated bars in T4 segments suggests the importance of a localized erosion and deposition processes in wider unconfined channels ( Figure 7). In this comparison, a lower percentage of barely vegetated bars in T4 segments indicates more resilient conditions or a capacity to absorb changes in water and sediment discharge without significant morphological response (Poff et al., 1997).
Future alternatives of the proposed approach could incorporate a spatio-temporal dimension to the process-based analysis by considering satellite images of multiple years or months (or even weeks depending on availability) allowing the assessment of single extraordinary events such as the disturbance generated by a flood or the alteration caused by a human impact such as a weir construction.
Moreover, the future availability of satellite images with higher resolution would facilitate a more detailed classification of landforms and mesoforms, together with a better characterization of several types of artificial interventions (i.e., artificial levees, dykes, lining works, transversal and crossing structures, etc.). As geospatial data are evolving fast to provide new products with higher resolution and shorter acquisition time intervals these shortcomings for the hydromorphological assessment at reach scale will be reduced.

| Towards an environmental assessment tool in ephemeral channels
The size, shape, and character of a river channel reflects components of prevailing flow and sediment bedload (Leopold et al., 1964). Thus, any change in sediment and water fluxes, occurring either naturally or as a result of human activity, implies an adjustment of the channel to new equilibrium conditions (Downs & Gregory, 2004;Kondolf, 1995).
The proposed stream classification does not inherently aim to guide geomorphic recovery, but it may offer a first-cut analysis to help identify and evaluate degraded segments based on anomalous morphosedimentary patterns. Open geospatial data currently available for most countries (Sentinel-2 images, LiDAR data) and GIS resources make the proposed approach suitable for regional scale baseline surveys of reach types and river landscape units. Based on the spatial configuration of morphological characteristics and stream types, it is possible to identify dominant river processes (entrenchment, sediment continuity, bank erosion, etc.) and their relation to human impacts and disturbances.
As an example, we select an obvious example of the impacts of transversal barriers such as dams on the measured morphosedimentary parameters and river type spatial configuration within the studied catchment. The morpho-sedimentary impacts at the Maria Cristina Reservoir (Figure 1) is patent in the upstream segment but also downstream. At the tail of this reservoir, channel slope decreases along with a reduction on sediment delivery and morphologic activity enhancing vegetation growth (segment V9 of T1; Figure 9a). Note that this T1 segment, typical of headwater zones, with a low channel slope and located next to the dam, is itself anomalous. Downstream from the dam wall, the location of stream channels T1 and T2 with narrow active channels and anomalous slopes suggests a sediment transport decay, which leads to adjustments in channel geometry (Figure 9a,b). The most degraded reaches occur in a short reach below the dam (V10 to V12) with a reduction of $50% in channel width in comparison to segments located upstream of the dam (Figure 9b). River adjustments continue downstream to the T4 (V11) and T3 (V13 and V14) segments, causing a reduction of active channel width and slope (Figure 9a,b).
Morphological changes in channel width are distinctive of reservoir impacts worldwide, with cases where channel width decreased by 90% or increased by as much as 100% (Downs & Gregory, 2004;Williams & Wolman, 1984). These observed geomorphic alterations therefore indicate a change in the sedimentary continuity and connectivity at this point (Fryirs, 2013;Wohl, 2017  and water flows (Gregory & Park, 1974).
Other segments suggesting sediment discontinuity are associated with in-stream gravel mining. In particular, V12 (T2) and V4 (T3) show narrowing and discontinuity of the active channel assumed to be related to channel incision caused by aggregate mining, based on evidence from field surveys and reach-scale photogrammetric mapping Calle et al., 2020). Other effects of mining activity include the formation of knickpoints that eventually may migrate several kilometres upstream during flooding (Kondolf, 1994). Field evidence of repeated knickpoints within a single segment is potentially detected by our spatial analysis from anomalous values of slope gradient at specific segments (e.g., V2, V7, V10; Figure 5b).

| CONCLUSIONS
An essential first step in the analysis of IRES status is to classify channel types representing the spatial diversity of geomorphic processes and their dynamic interaction with vegetation. In this study, we present an objective, semi-automatic and replicable approach for the segmentation and classification of channels with similar morphology using (1) automatic segmentation procedures using GIS tools and statistical methods and (2) reach-scale landform classes using machine learning algorithms applied to satellite images. Four channel types were defined from a cluster analysis of 62 segments of the Rambla de la Viuda catchment (1500 km 2 ), with the typology based on differences in elements of the valley/channel geometry, fluvial landform assemblage, and hydraulic conditions. Segments T1 occurs on narrowly confined segments in headwater areas with a high stream gradient, reduced riparian zones, and boulder dominated bedforms (cascade or step-pool channels). Segments T2 occurs in partly confined segments in piedmont sectors with a high-moderate gradient, narrow riparian zones, and discontinuous bed gravel landforms (plane bed to step-pool). Segments T3 comprises semiconfined reaches in lowlands with a moderate gradient, broader riparian zones, and continuous bed gravel landforms (lateral gravel bars). Finally, segments T4 consists of unconfined segments with a low channel gradient, welldeveloped riparian zones, and wide continuous active gravel bed landforms (multithread channels).
F I G U R E 9 Diagrams of watershed, topographic variables and stream type classification. (a) Drainage area versus channel slope for the identified segments with indication of the classification type. (b) Drainage area versus active channel width. The arrow illustrates the generalized relative trend in sediment load broken at the Maria Cristina dam barrier. Numbers of segments refer to Figure 5 The applied catchment-scale stream analysis and channel classification is an efficient tool for: (1) understanding the geological, structural, and vegetation controls on the spatial distribution of channel morphology; (2) interpreting fundamental processes (depositional and erosional) and disturbance regimes at the reach scale through links between alluvial landforms and vegetation density; (3) identifying the location of reaches that supply and store sediments; (4) predicting, based on the former, the sediment continuity through the river network; and (5) identifying geomorphic disequilibrium and its causal links with natural and anthropogenic impacts.
The applied methodology is particularly well-suited for ephemeral streams whose environmental quality assessment depends on geomorphological conditions that support riparian habitats. It is important to recognize that channel categories and the distribution of geomorphic processes (process domains) depend upon internal controls (e.g., climate, geology and topography), therefore, the classification presented here is not intended to be universal in its channel types but in its methodological approach. A fundamental advantage of this approach is the objective delineation of stream segments along a continuous river network, which enables systematic characterization of extensive Mediterranean regions. Currently, a limiting factor in fluvial landform delineation is the spatial resolution of the available open-access satellite imagery, but it is likely that this drawback will be overcome by new products released by space agencies (e.g., ESA, NASA). Automated approaches that are applicable over large regions lay the foundations for an essential characterization of river channels, including their management, conservation and restoration of river corridors. In particular, the application in IRES is highly appropriated because of the lack of summer flows, which leaves the bed exposed, and the limited riparian canopy relative to perennial river channels.