Modeling soil erosion and deposition within a Mediterranean mountainous environment utilizing remote sensing and GIS-Wadi Tlata, Morocco

Nutzungsbedingungen Die ETH-Bibliothek ist Anbieterin der digitalisierten Zeitschriften. Sie besitzt keine Urheberrechte an den Inhalten der Zeitschriften. Die Rechte liegen in der Regel bei den Herausgebern. Die auf der Plattform e-periodica veröffentlichten Dokumente stehen für nicht-kommerzielle Zwecke in Lehre und Forschung sowie für die private Nutzung frei zur Verfügung. Einzelne Dateien oder Ausdrucke aus diesem Angebot können zusammen mit diesen Nutzungsbedingungen und den korrekten Herkunftsbezeichnungen weitergegeben werden. Das Veröffentlichen von Bildern in Printund Online-Publikationen ist nur mit vorheriger Genehmigung der Rechteinhaber erlaubt. Die systematische Speicherung von Teilen des elektronischen Angebots auf anderen Servern bedarf ebenfalls des schriftlichen Einverständnisses der Rechteinhaber.


Introduction
The removal of permanent Vegetation has accelerated erosion and deposition in many stream basins within the Mediterranean region. Accelerated erosion, resulting largely from overland flows (Trlbak 2000), along with deposition increases in low lying sinks, especially reservoirs, results in land degradation. The tributary steam basin of the Wadi Tlata in northeastern Morocco exemplifies these phenomena. Due to agricultural and grazing activities which altered the area's Vegetation cover, accelerated erosion and deposition is being experienced throughout the catchment. The spatial alteration of the wadi's erosional/depositional balances has clear impacts on the economic potential of the area. This study's aim is to gain insights applicable in assessing land degradation occurring in similar Mediterranean environments. In addition, this investigation's approach ülustrates how a relatively quick assessment of an area's erosional and depositional properties can be determined even in areas with minimal data. The investigation utilizes space imagery over a 15 year period (1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002) and analyzes these data within a geographical information System (GIS). Based on the outcomes of the analysis, an ordinal ranking of the spatial properties of erosion and deposition is obtained. This ranking can be used to determine critical areas within the catchment needing implementation of conservation practices to curtail land degradation. 2 Background and method 2.1 Study area The 123 km2 of the Wadi Tlata Catchment, located in northeastern Morocco, has a generally northwest to southeast orientation (Fig. 1). The basin has a rugged topography, elevations ränge from 1,373 in the northwest to 560 meters at the stream's mouth. The dominant underlying bedrock, tertiary marls (Leblanc 1979), produces soüs easily eroded when exposed by human activities. Given the area's relief and climate, occasional intense winter rainfall and a paucity of precipitation during the hot summer months, when Vegetation is disturbed, the potential for accelerated erosion is high. These physical attributes, along with an average population density of 76/km2 and ubiquitous agricultural activities, result in slopes largely denuded of their natural ground cover.

Methodology
This study utilizes space imagery and field data to determine the topography, land use, and ground cover. These data are then processed and the spatial distribution of soil loss and Sedimentation are estimated using a RUSLE (Lewis et al. 2005) and Sedimentation (flow) model within a GIS framework (Fig 2). The extraction of the information sets (e.g. land use) utilized topographic maps, field work, and satellite imagery. The development of the digital elevation model (DEM) (derived from a topographic map with 10 m contours) and digjtization of the data layers were created using Are View. The analysis of these data sets, including image processing and the modelüng of soil loss and Sedimentation utiüzed Idrisi (Fig. 2).
Most prior studies concerning soil loss based on the USLE/RUSLE (Universal Soil Loss Equation/Revised Universal Soil Loss Equation) model developed by Wischmeier and Smith (1978) and refined by Renard et al. (1991) do not examine the related process of deposition (Bonn 1998;Dialo 2000;Renard et al. 1997;Wischmeier 1976) since RUSLE's conceptualization only permits soil loss to be estimated. Some authors have asserted the possibility of identifying deposition zones using accurate DEMs (Mitasova et al. 1996; Moore & Burch 1986). Previously, Van Rompaey et al. (2001) integrated soil loss and deposition into a Single model based on RUSLE. However, their model requires the Sediment delivery ratio which varies within catchments and is rarely known for most areas. Others (Fernandez et al. 2003 andMutua et al. 2006) have extended RUSLE to also estimate Sediment yield. The approach presented here directly integrates RUSLE's Output data into a flow model which allows net soü loss and deposition to be accessed within a GIS framework without additional data requirements. This represents an extension of the incorporation of RUSLE into a GIS framework by Lewis et al. (2005).
Space imagery analysis was used to determine the two factors considered to strongly affect erosion and deposition: land use and land cover. Four images were analyzed (Spot HRV 15 Nov. 1994  A four step procedure was undertaken to obtain data on land cover and land use from the satelüte imagery. First, the Visual interpretation aüowed identification of broad classes of land cover. Second, the broad «Visual» groupings were refined by an unsupervised Classification based on theü spectral signatures.Third, post-classification filtering was undertaken to remove ükely errors from the data. Finaüy a temporal analysis was initiated to determine the land cover and land use changes that took place during the study period. The land cover/land use data were inputs into the RUSLE model. The Output from this model and the SEDIMENTATION module permitted the examination of spatial and temporal patterns and gave insights into soü loss and Sedimentation totals.
A. Visual interpretation. Forest areas were delimited based on the satellite images' color composition. The forest areas were then classified as «natural protected» or «afforested» based on previous field Visits. Differentiation between «bare ground areas» and «badland zones» was made in a similar manner. Whereas the Classification of forest, bare ground and badland areas is considered to have a low error ratio, cultivated lands are less certain. It was hard to differentiate the various types of cultivated land. Only where olive tree plantations were visible was error probability considered small in identifying the cultivated land type.
B. Unsupervised Classification. In the next step, an unsupervised Classification of the visually determined groupings was undertaken. Within each Visual grouping, pixels were grouped on the basis of similar spectral signatures (Iain et al. 1999). Ten classes of specialized land cover, based on the known characteristics of the Visual Classification were chosen in this initial grouping. Next, several iterations of this initial grouping were performed during which arrangements and groupings of the classes were refined to create more homogeneous and better differentiated groups. This process results in classes that reflect real land use. To improve these results.it was necessary to define several subclasses, with slightly different spectral properties. For example, the spectral properties of identical forest It was possible to identify two subclasses (forest on the slope exposed to the south and forest on shaded slopes). Consequently, by specifying the spectral definition of each corresponding class, Classification quality was appreciably improved. Using this logjc it was possible to identify the foüowing classes: badlands, arable lands, and permanent Vegetation (protected natural forest and reforestation).
The classes were then refined to decrease within variability. In certain cases, where land Covers are known, specific land cover can be isolated by carrying out a simple thresholding of the numerical values in either one (monospectral) or several (multispectral) spectral bands.A class of objeets is defined by a ränge of values (i.e. all pixels with numerical values that correlate with a given known ground cover on a specific image).
Particular use was made of the «red» TM3 band as it allows for discrimination between plants species. For example, the higher TM3 band values were found to represent protected forest cover. These results along with utilization of the Normalized Difference Vegetation Index (NDVI), allow differentiation of land cover/land use into six classes: badland/degraded, annual cereal crop lands, arable lands with trees, oüve trees, reforested areas, and natural forest areas.
C. Post-classification filtering. Finally, to develop the tables of land use and ground cover, it was necessary to refine the results obtained from the previous Classification procedures. For this purpose, various filtering techniques were implemented (Fig. 3). Attention was given to Single isolated pixels, pixels that were not classified, and pixels that were clearly misclassified. These problem areas often oecurred along boundary conditions including two different Vegetation Covers or land use classes as well as along the catchment divide.
Single isolated pixels were treated as errors as land use units in the study area are not this small. To homogenize isolated pixels a filtering technique was used to remove all one and two pixel islands and most three pixel islands. The only three pixel islands that the filtering technique did not remove were those clustered in a corner shaped configuration (Fig. 3a). The technique used to remove these isolated pixel islands was a 3-by-3 moving window. By moving the window pixel-by-pixel across the entire image surface, all islands could be located. Location of isolated pixels was aecomplished through a two-stage sequential filtering process. The first stage exploits a relaxed filtering approach with the isolation threshold set at 87.5%. This means a pixel with a certain class will be considered isolated only if it is surrounded fully by any other class or surrounded by a Single dominant class which aecounts for 87.5% of the neighboring pixels, i.e. 7 out of 8 in a 3-by-3 neighborhood (Fig. 3b). Once an isolated pixel is identified, it is merged into the dominant neighboring class by Converting its original class value to the value of the dominant class (Fig. 3c). The second filtering step utilizes an isolation threshold of 100%. This step deals with central pixels that are completely surrounded by a different, but homogenous class (the Situation of Single pixels being surrounded by two or more classes were treated in step 1) (Fig. 3d). D. Combination ofimages: The multi-temporal analysis.
One principal advantage of utüizing satelüte images over an extended time frame is that ground cover and Vegetation changes over time can be deciphered. The approach presented in this paper permitted a reüable ordinal ranking of soü loss and deposition for the study area, a location with a minimal environmental data base. Four land use/ land cover maps (for each satellite image) for the 1987 to 2002 period were developed. These land cover/land use data are inputs into the RUSLE and SEDIMENTATION modules. The Outputs from these modeis permit the examination of the spatial and temporal patterns and the soü loss and Sedimentation estimates.
3 Modelling soil loss and deposition
The study's R value (56.3) was based on the formula of Rango & Arnoldus (1987) and derived using 1977-1994 data from the Arn Boukaüa weather Station situated within the Wadi Tlata. K values were derived using the Wischmeier nomogram (Wischmeier & Smith 1978), a pedologjcal map and the area's infiltration properties (Trlbak et al. 2006). The K factor for the river basin's soüs ranged from 0.22 to 0.46. The LS factor was computed using GIS Software (Lewis et al. 2005). The crop cover factor (C) was calculated using the satelüte images, as weü as ground truth observations.The values assigned to land cover ranged between 0.08 for natural protected forest to 0.75 for poorly vegetated badlands. These values are based on the previous calibration of this factor for the local agricultural System (Kalman 1967). The P factor value of 1 (no conservation) was assigned for the whole area since üttle conservation is practiced in the study area, and where practiced, it is poorly implemented.
3.2 Estimation of soil loss After determining the RUSLE input parameters, soil loss was estimated using the RUSLE module (Lewis et al. 2005). The execution of the module requires thresholds for slope, azimuth and slope length.The maximum threshold slope length is the distance beyond which overland flow becomes concentrated. RUSLE recommends lengths less than 122 meters (Renard et al. 1997). For the approach presented here, thresholds of 5 degrees for slope, 10 degrees for azimuth and 100 m for maximum slope length were utüized. These values are reasonable for the topographic characteristics of the wadi.

Estimating net soil loss and deposition
One limitation of RUSLE's conceptualization is that its application to any land unit must result in soil loss; it cannot predict deposition. Yet clearly within many agricultural fields and river basins, deposition oecurs.
The reason RUSLE only estimates soil loss is that it views soil loss in a static manner. Soü loss is estimated for all areas based on their R, K, L, S, C, P parameters.
The downslope movement of soil loss from one area (patch) to another is not conceptualized within RUSLE. By incorporating the RUSLE algorithm into a GIS framework, it is possible to examine the spatial down slope movement of soil loss and from these movements determine net soil loss and deposition. Net erosion or deposition for a gjven area is determined in the foüowing manner when applying the SEDIMENTATION module Software: determination of patch soil loss, establishment of direction of soil movement and calculation of proportional soü loss.
First, soil loss for each patch (a construeted homogeneous polygon with regard to slope aspect, slope, land cover, conservation practice, erodability, and erosivity) within the river basin is determined (Lewis et al. 2005). Next, the average elevation for each patch is identified. This is derived by averaging the highest and lowest elevations in each patch. Then the river basin's highest elevation is located. The direction of soil movement is estabüshed by taking the relative elevation differences between contiguous patches into account. The amount of soü loss that moves into the surrounding lower patches is assumed to be proportional to the length of the common boundary between the higher patch and the lower patches (Eastman 2006). Next, the net soil loss and deposition in all the lower patches is calculated.
The proportional RUSLE soil loss moving from the higher patch is compared to the RUSLE soü loss derived for the lower patch. For example, the proportional soil loss of Patch A (the higher one) to lower patch B is 1 Mg yr4. Patch B's initial RUSLE computed soil loss is 3 Mg yr1. Thus, the net soil loss of Patch B is calculated by the SEDIMENTATION module as the difference between the amount of Sediment trans-ported into the patch and the patch's RUSLE soil loss value. In the previous example, 3 Mg yr4 (B) -1 Mg yr1 (A) results in a net soil loss for Patch B of 2 Mg yr4 (Fig. 4a). Similarly, if Patch C located downslope from Patch B, has an estimated soü loss of 2 Mg yr4, its net deposition wiü be 1 Mg yr4 (Fig. 4b).
4 Results and discussion 4.1 Land use and land cover changes The multi-temporal analysis results of the four sateüite images (1987-2002) using Visual comparison and cross tab analysis revealed the foüowing changes in the six land use and land cover classes (Table 1): 1: badlands/degraded lands increased in areal extent in each time period; 2: cultivated areas of largely cereal crops decrease in areal extent in each time period; 3: lands under olive eultivation generally increased; 4: reforested land, which are lands that have degraded too much for eultivation, increase slightly; 5: mixed farmlands (comprised of cereal crops and tree crops) showed an increasing trend over the time period; and 6: areas of natural forest remained relatively stable.
Thus, reforested areas and areas with natural shrub forest were relatively stable in terms of cover during the 15 year period; badlands and degraded lands almost doubled in areal extent and tree crops expanded while areas solely farmed in cereals declined. Nevertheless, farmlands in cereal produetion remained the dominant land use.

Soil loss estimates
After land use was determined, estimates of average annual soil losses for each ground cover were computed using RUSLE. These results for each image are presented in The spatial patterns of soil loss within the river basin did not differ greatly over the 15 year study period. Given this finding, the average soil loss values for each land use are used to iüustrate the spatial pattern of estimated soil loss for the catchment (Fig. 5). Overall the lowest soil loss values are found on the gentlest values largely reflect the combination of land cover and slope. Soil erodability properties have a slight impact. Neither the rainfall nor conservation factors contribute to the spatial pattern as the rainfall factor varies only slightly in the basin and conservation is not practiced throughout the wadi. Applying the SEDIMENTATION module to RUSLE aüows for net soil loss for each patch to be estimated.
This approach should result in a more accurate assessment of net soil loss. Once downslope movement of soil is incorporated into the estimates, soil losses are significantly less (Table 3). Of particular interest is that three classes of land cover (annual crops, mixed farmlands, and reforestation) based on average values fiip from an erosional mode into a depositional dynamic. In addition, areas under olive tree eultivation have their soü losses reduced from an unsustainable estimated (RUSLE) average soil loss of 77 Mg ha4yr4 to 0.5 Mg ha4yr4. Both the badland and natural open forest classes remain in the erosion mode. The open forests and badland classes with generally steep slopes minimize depositional potential.
Examination of the Standard deviations of the net soil losses (Table 3) give further insights into the soil dynamics oecurring in the catchment. The land use classes of annual crop eultivation and mixed farmlands oecupy a large proportion of the river basin. In the areas where these eultivations occur on moderate and steep slopes, the data indicate these are zones of excessive erosion. But when these farming activities occur on the plains and Valley bottoms, they experience net deposition. The net soil loss map (Fig. 6) shows that in the eastern and southeastern portions of the basin, areas where highly erodible soüs derived from marls and sandy marls are found (see Tribek 2000 for details of soil properties), net soil losses of 50 Mg ha4yr4 occur. Given that these zones oecupy approximately 1150 ha or 9.5% of the basin, the soil loss data indicate these areas must be degrading and are likely to experience declining yields in the near term or eventually go out of produetion unless interventions occur shortly.

Discussion
A comparison of the results with previous Undings (Al Karkouri 2003;Dhman 1994;El Garouani et al. 2003;Heusch 1970;Rahhou 1999) highlights the dampening effect that a dual focus on erosion and deposition has on catchment erosion risk assessment. An overall decrease oecurs in the erosion rates of a river basin. Although the studies mentioned above agree that a high erosion risk exists in mountainous areas in Mediterranean regions, this paper indicates that this could only be confirmed for the eastern sections of Wadi Tlata.
The multi-temporal analysis of land cover and land use indicates that the erosion risk has increased on moderate to steep slopes in the eastern portion of Wadi It should be emphasized that the results presented approximation of reality, despite the likelihood that in this analysis only provide a relative spatial and the contribution of the latter is minor (El Garouani have not been incorporated into the assessment. For example, the presence of stones on the surface in many areas was not included despite the likelihood that they could affect erosion rates. Further, the erosivity factor (R) was based on annual precipitation monitored near the mouth of the basin. The R values used for the whole basin were extrapolated from that point based on the basin's precipitation patterns. Actual measurements of R would improve the reliability of this factor.
In spite of these deficiencies, the method presented in this study produced vaüd insight into both the temporal and spatial patterns of erosion, erosion risk and Sedimentation in the basin. With this approach, it was possible to identify zones which, in the long-term, are likely to experience excessively high soil losses and deposition.

Conclusion
This article presents a rapid approach to assessment of spatial and temporal patterns of soil loss and deposition utilizing remote sensing and GIS within the context of a stream basin in a Mediterranean regjon. It reports on the extension of the RUSLE model to permit net soil loss and depositional areas to be identified and assessed. This is considered of particular relevance where soil losses are transported beyond field boundaries. The results indicate that former studies in the area, which were neither sensitive to deposition nor concerned with the movement of the soil loss through the area probably overestimated erosion in many areas, especially on the gentle slopes. In particular, deposition occurring along the Valley bottoms was not identified.
The approach used aüows for estimation of net soü loss/ deposition for any area within a catchment as weü as for assessment of their spatial patterns. As the approach considers both soü loss and deposition, the results approximate reaüty to a greater degree than those determined by RUSLE alone. Although the accuracy of the soü loss and deposition values derived from the combined RUSLE/SEDIMENTATION module approach is open to discussion due to data limitations of the area, the method presented aüows at the very least a rapid assessment of soü loss and deposition and a relative ranking of erosion and deposition within a study area.
With a basic data base such as that avaüable for the study area and typical for many areas within developing countries, results gained from this approach can allow decision makers and developers to obtain a relative picture of existing conditions in any area and simulate land use/land cover scenarios in order to plan interventions to curtaü erosion and deposition.   (1987,1994,2000,2002). The purpose of the study was to establish a means of rapidly determining land cover and land use changes, as well as spatial patterns of erosion and deposition, in areas with relatively poor data bases and where soil loss results primarily from nonchannelized flows. The procedure selected involved the foüowing: establishment of land use class distribution and size for each year of Observation, static estimation of soü loss, calculation of net erosion and deposition, and prioritisation of critical areas. Thus, for the targeted 123 km2 Tlata catchment of northeastern Morocco, six main land use classes could be defined (highly degraded lands, annual cereal crops fields, mixed farmlands, olive trees, reforested areas, and natural protected forest). Analysis of remote sensing data allowed estabüshment of the areal distribution of each land use class for each year. Soil loss was estimated using a RUSLE module integrated in a GIS framework. These static areal estimates of soil loss were then fed into a Sedimentation algorithm that modeis downslope movement of soil loss. From the resulting spatial (flow) movements, net erosion and deposition for each time period could be estimated. The results permit, at the least, an ordinal ranking of erosion and deposition within the basin. This Supports decision-making processes on prioritization of areas where interventions are needed to ameliorate or prevent land degradation.