- Email: [email protected]
Effect of vegetation cover on soil erosion in a mountainous watershed
Catena 75 (2008) 319–325
Contents lists available at ScienceDirect
Catena j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / c a t e n a
Effect of vegetation cover on soil erosion in a mountainous watershed P. Zhou a,⁎, O. Luukkanen b, T. Tokola c,d, J. Nieminen c a
College of Resources and Environment, Northwest Agriculture and Forestry University. P. O. Box 81, 712100, Shaanxi, China Viikki Tropical Resources Institute, Department of Forest Ecology, University of Helsinki, P. O. Box 27, FIN-00014, Helsinki, Finland c Department of Forest Resource Management, University of Helsinki, P. O. Box 27, FIN-00014, Helsinki, Finland d Faculty of Forestry, University of Joensuu, P. O. Box 111, FIN-80101, Joensuu, Finland b
a r t i c l e
i n f o
Article history: Received 3 February 2008 Received in revised form 17 June 2008 Accepted 28 July 2008 Keywords: k-NN Restoration RUSLE Soil erosion Upper Min River watershed Vegetation cover
a b s t r a c t We applied the Revised Soil Loss Equation (RUSLE) to assess levels of soil loss in a Geographic Information System (GIS). In this study, we used the k-NN technique to estimate vegetation cover by integrating Landsat ETM+ scenes and field data with optimal parameters. We evaluated the root mean square errors and significance of biases at the pixel level in order to determine the optimal parameters. The accuracy of vegetation cover estimation by the k-NN technique was compared to that predicted by a regression function using Landsat ETM+ bands and field measurements as well as to that predicted by the Normalized Difference Vegetation Index (NDVI). We used a regression equation to calculate the cover management (C) factor of the RUSLE from vegetation cover data. On the basis of the quantitative model of soil erosion, we explored the relationship between soil loss and its influencing factors, and identified areas at high erosion risk. The results showed that the k-NN method can predict vegetation cover more accurately for image pixels at the landscape level than can the other two methods examined in this study. Of those factors, the C-factor is one of the most important affecting soil erosion in the region. Scenarios with different vegetation cover on high-risk areas showed that greater vegetation cover can considerably reduce the loss of soil erosion. The k-NN technique provides a new method to estimate the C-factor for RUSLE erosion mapping. The quantitative model of different vegetation cover scenarios provides information on how vegetation restoration could reduce erosion. © 2008 Elsevier B.V. All rights reserved.
1. Introduction Soil erosion by water is a worldwide environmental problem which degrades soil productivity and water quality, causes sedimentation and increases the probability of flood. The Upper Min River (UMR) watershed in the Upper Yangtze Basin is an environmentally fragile area due to deforestation and soil erosion. Up to 44% of the land has been described as degraded in the area (Ye et al., 2003; Wu et al., 2003). Different methods have been developed to detect eroded areas and to assess erosion loss. For instance, the qualitative classification of soil erosion can be assessed by FAO (2006), and large and middle-size eroded areas can be directly identified with Landsat and SPOT imagery (Langran, 1983; Millington and Townshend, 1984); the quantitative soil loss can be modeled with the universal soil loss equation (USLE) (Wischmeier and Smith, 1978), its revised version (RUSLE) (Renard et al., 1997), the Soil Erosion Model for Mediterranean regions (SEMMED) (De Jong,1994), 137CS techniques (Zhang et al., 2003), and the Water Erosion Prediction Project (WEPP) hill slope model (Grønsten and Lundekvam, 2006). Of these models, the USLE and RUSLE are the most widely used, providing a convenient tool for soil loss evaluation by taking into consideration rainfall, ⁎ Corresponding author. Tel.: +86 29 87080321; fax: +86 29 87080055. E-mail address: [email protected] (P. Zhou). 0341-8162/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.catena.2008.07.010
topography, conservation support practice, soil, and vegetation. Although they are empirical models for assessing long-term averages of sheet and rill erosion based on plot data, the calculation of their factors can be improved and adapted to enable their application to various spatial scales and region sizes in different environments using GIS (Warren et al., 1989). For better prediction of soil loss in complex terrain, the modelling of dispersed flow over the surface can improve the slope length and steepness factor (LS) (Tarboton, 1997). Of these six RUSLE factors, the cover management factor (C-factor) is an important one affecting soil erosion in a given region. However, modelling the C-factor for complex terrain is difficult. The C-factor in the soil loss equation is defined as the ratio of soil loss from land cropped under specified conditions to the corresponding loss from clean-tilled, continuous fallow (Wischmeier and Smith, 1978). Land use classification is often used to map vegetation types that differ in their effectiveness to protect the soil. After classification, C-factors are assigned according to a qualitative ranking of vegetation types (Wischmeier and Smith, 1978; Morgan, 1995). However, the proportional vegetation cover varied widely even for the same vegetation type. The direct application of the C-factor from the RUSLE is based on prior land use (PLU), canopy cover (CC), surface cover (SC), surface roughness (SR), and soil moisture (SM) (Renard et al., 1997). However, assessing PLU, CC, SC, SR, and SM simultaneously for a large area not covered mainly by agricultural lands is difficult. The Normalized Difference Vegetation Index (NDVI), defined
320
P. Zhou et al. / Catena 75 (2008) 319–325
as the near infrared reflection minus the red reflection divided by the sum of the two (Tucker, 1979), was calculated from satellite images for C-factor estimation (Thiam, 2003; Wu et al., 2004) because green vegetation is more reflective in the near infrared part and less reflective in the red part of the spectrum. However, De Jong (1994) found that the relationship between Landsat-derived spectral indices and vegetation attributes was quite poor for Mediterranean France. The non-parametric k-nearest neighbour (k-NN) technique, widely used in a variety of forest estimation and biomass mapping applications over the years (Tokola, 2000; Franco-Lopez et al., 2001; Katila and Tomppo, 2001), provides a new method to map proportional vegetation cover. The k-NN technique is a non-parametric approach to predicting values of point variables on the basis of similarity in a covariate space between one point and other points with observed values for the variables (Tomppo, 1991). Using the k-NN method, we cannot only predict variables such as vegetation cover, aboveground biomass, and stand volume or parameters for each pixel, but also examine the covariance structure of field variables and produce pixel-by-pixel maps. In this study, we used the k-NN method to estimate vegetation cover data and produced a pixel-by-pixel map by integrating the satellite images and field data with optimal parameters. The major objectives of this study were to produce a vegetation cover map using the k-NN method and testing the performance of this method, to quantitatively evaluate soil erosion at 25-m resolution for a 7400 km2 area of the watershed using the RUSLE model with predicted factors, especially an improved C-factor, to test the relationship between soil erosion and its influencing factors, as well as to model soil loss risk under different vegetation cover scenarios.
2. Study area The study area is located in the Upper Min River (UMR) watershed (Fig. 1), with an area of 23,040 km2. The target watershed is in Sichuan province, Southwest China, between 31°–33° N and 102°–104° E. The Min River is also one of the most important tributaries of the Upper Yangtze River. The climate is governed by the northeast and southwest monsoons. A complex topography, with elevation ranges from 600 m to more than 6000 m a.s.l. (above sea level), results in steep gradients of rainfall on both spatial and temporal scale. The annual precipitation ranges from 405 mm to 1950 mm in different parts of the watershed, and around half of the precipitation falls in July, August and September (Zhou, 2008). The upper reaches of the Yangtze River are the region suffering from serious water erosion in China. Around 1.56 billion tons of soil is eroded in this region each year (MWR, 1999). The UMR watershed is fragile to water erosion due to steep terrain and deforestation. The forest cover had declined from 50% to 30% by 1950 and to 18.8% by the 1980s, and to only 5%–7% along the main river (Wu et al., 2003). At present, around 21% of the entire watershed has forest cover. The vegetation here ranged from subtropical evergreen broadleaved forest to alpine meadows. Our 625 inventory plots were randomly placed in the middle and upper reaches of the UMR watershed (Fig. 1) over an area of about 7400 km2. The vegetation ranges from subtropical evergreen broadleaved forest to alpine meadows. Slope angles range from 0 to 77.2° with a mean of 25.9° (Zhou et al., 2006). The area comprises three main soil orders (Alfisol, Semi-alfisol, and Semi-aquatic) and faces
Fig. 1. Location of the research area and the sample plots.
P. Zhou et al. / Catena 75 (2008) 319–325
many environmental problems such as land and soil degradation, soil erosion, mud- and landslides, earthquakes, and the growing desertification of some valleys along the main Min River. 3. Materials and methods 3.1. Soil erosion by RUSLE The RUSLE has served to evaluate and predict soil erosion in the region on a pixel-by-pixel basis. The soil loss (A) due to water erosion per unit area per year (Mg ha yr− 1) was quantified using the RUSLE with the following equation: A ¼ R K LS C P
ð1Þ
where A is the average soil loss due to water erosion, R the rainfall and runoff erosivity factor (MJ mm ha− 1 h− 1 yr− 1), K the soil erodibility factor (Mg h MJ− 1 mm− 1), L the slope length factor (contributing area), S the slope steepness factor, C the cover and management practice factor, and P the support practice. We generated a digital elevation model (DEM) with a 25-m cell size grid using both contour lines and stream data to ensure hydrological accuracy. The data used to generate R, LS, and K included precipitation data, DEM, satellite images, inventory data, digitised stream data, soil map and soil erodibility data. Zhou et al. (2006) have separately described the method of raster maps generation with regard to R, LS, and K factors. Precipitation data measured with rain gauge from 1998 to 2002 was collected from 40 meteorological stations within and surrounding the research area. A multivariate geostatistic cokriging model was used to produce the raster map of R-factor. We used the standard for soil erosion intensity issued by the Ministry of Water Resources, China (MWR, 1997) to classify soil loss in this specific watershed. The standard was determined mainly based on field experiments in China and was guaranteed to be applicable for the whole China. According to the predicted soil loss amount, the potential soil loss was divided into six ordinal classes. We then calculated the area and percentage of each soil loss class by generating zonal statistics in ArcGIS. 3.2. C-factor estimation We employed the k-NN technique to predict vegetation cover for image pixels and used consecutive Landsat ETM+ scenes WRS2 130/ 037 and 130/038 from 10 July 2002 (GLCF, 2002) to estimate vegetation cover. To ensure compatibility between the images and the ground data, each image was rectified and georeferenced to the Universal Transverse Mercator system (UTM): WGS_1984_UTM_Zone_48N. The vegetation cover image was constructed using ETM+ bands 1, 2, 3, 4, 5 and 7. We recorded vegetation covers at three levels (canopy cover, under canopy cover, and overall vegetation cover) from 625 sample plots and measured the canopy cover with a spherical densiometre (Janey and Block, 1994). The field inventory was conducted in June, July and August 2004 using strata-delineated sampling with the aid of the global positioning system (GPS). Besides the measurement on vegetation cover at three levels, other recordings were also made such as the elevation, slope, aspect, forest type, soil type, plot condition, land use. More than 5000 field pictures were taken from the field and later on the pictures have been also used for visually checking the results. We calculated the C-factor of the RUSLE from the predicted vegetation cover image using a regression function Eq. (2). The equation was build by using over 200 soil loss ratios measured from 30 runofferosion plots under both natural and simulated rainfall events in the Three Gorges area of the Yangtze River basin (Yang and Shi, 1994). C ¼ 0:6508−0:343 log c 0 < c < 78:3%
ð2Þ
321
3.3. The k-NN technique for vegetation cover estimation The data for the k-NN method used in this study included a georeferenced satellite image and field plot data containing the following values: the x and y coordinate locations of plots, their corresponding satellite image spectral values from six bands, and total vegetation cover. To predict the vegetation cover using k-NN technique, consider an ETM+ pixel to be a point, let Y denote vegetation cover. The estimation of variable Y for pixel j can be computed from Eq. (3). Yˆ j ¼
! !−1 k k 1 ∑ wji ∑ wji Yji k i¼1 i¼1
ð3Þ
where k is a predetermined constant, 1 ≤ k ≤ n, {wji} are the weights of the points to be selected, and Yji = 1 ,…, k, is the value of variable Y in the sample plot i corresponding to the pixel j, which is the ith closest pixel in the spectral space to the pixel j. The geographical distance ranging from 1 km to 100 km to the potential nearest neighbours has been tested. With the leave-one-out method, a k-NN prediction (Yˆi) is sequentially obtained for each Yi,i = 1 ,… k, while Yi itself are not included in the mean forming of its own k-NN prediction. The quality of the predictions can be evaluated by comparing the observations Yji = 1 ,…, k and the corresponding predictions Yˆj. Specific parameters for the k-NN technique must be estimated before implementation. The particular spectral bands should be selected to calculate distances dji between spectral band Xj and each element of the set (Xi,i = 1,… ,n). Distances between neighbours were computed using weighted Euclidean distance metrics Eq. (4). dji ¼
M 2 ∑ vm Xjm −Xim
1=2
m¼1
ð4Þ
where {vm} are variable weights for the feature m, and M are the selected bands. Variable weights are used in the equation because not all the features in the feature space share the same degree of influence on the prediction of vegetation cover for a given pixel. Finally {wji} must be tested. Three different weighting functions such as equal weights , inverse distance weighting, and inverse to the square of the distance weighting were used to compute {wji} .These weights are computed by choosing t = 0, 1, or 2 respectively Eq. (5). wji ¼ dji−t
ð5Þ
The accuracy of the estimate on vegetation cover can be examined using the root mean square error of cross-validation (RMSECV): " RMSECV ¼
1 n ∑ n i¼1
2 #1=2 −i ˆ Yi − Y i
ð6Þ
where Yˆi− i is the predicted value of the ith observation using the leaveone-out method without considering observation i. The optimal number of k as well as of the number of bands and weights were chosen when RMSE and bias were minimal. We compared the predictive performance of the k-NN method to that of a multivariate exponential regression model, as well as to that of the NDVI. The method explained by Hall et al. (2006), which tests have shown to yield the best model performance for predicting height and crown closure, was used to construct the multivariate model (Y = exp (B0 + B1X1 + … + BnXn)). Data used for building the multivariate exponential regression model in this study were satellite image spectral values from six bands, and field recordings on vegetation cover with GPS coordinates. The NDVI was computed from spectral data from red band 3 (RED) (wavelength 0.63–0.69 µm) and near infrared band 4 (NIR) (wavelength 0.75–0.90 µm) from the Landsat ETM+ (NDVI = (NIR − RED)/(NIR + RED)) (Tucker, 1979). A regression
322
P. Zhou et al. / Catena 75 (2008) 319–325 Table 2 Paired-sample t-test on observed and predicted vegetation cover using k-NN, multivariate regression and NDVI Methods
Paired differences (%) Mean
Std. Std. error 95% deviation confidence interval Lower
k-NN −0.608 11.472 Multivariate 0.035 15.054 regression NDVI 0.0002 15.465
Fig. 2. Vegetation cover map at 25 × 25 m pixel size predicted with k-NN method.
function was then built between calculated NDVI and measured vegetation cover for those sample plots. The paired-sample t-test (SPSS 12.0.1) was used to check the accuracy of the estimate by comparing the observed and predicted vegetation cover with these three methods. 3.4. Scenarios of vegetation cover on soil erosion Through a series of computations using map algebra based on Eqs. (1) and (2), we generated maps to show what effects could result from different cover management factors using ArcGIS software. First of all, soil loss intensities were compared under the current vegetation cover as well as under scenarios of no vegetation protection (C = 1) and good vegetation protection (C = 0.001). Five different vegetation cover restoration scenarios, such as a restoration of vegetation cover of less than 40%, 50%, 60%, 70%, or 78.3% for each respective value, were simulated to calculate how different vegetation cover affected the areas at high-risk for soil loss. Zonal statistics were used to calculate the proportions of different soil erosion intensities under each scenario. 4. Results 4.1. Vegetation cover map Vegetation cover with a 25 × 25-m pixel size was produced using the non-parametric k-nearest neighbour (k-NN) multi-source estimation method by integrating the satellite images and field data with optimal parameters (Fig. 2). The set of parameters chosen for the k-NN method to predict the vegetation cover included the image bands, the weight for each band, the distance, the number of nearest neighbours, the value of k, and the geographical reference area of the nearest field plots (Table 1). The leave-one-out cross-validation method was applied to calculate the RMSE and the average biases of predictions at the single-pixel level for different combinations of the k-NN
Table 1 Tested parameters for vegetation cover estimation using k-NN method Parameters
Values
K Geographical distance (km)
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 1, 5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95, 100 1, 2, 3, 4, 5, 7 0.7, 0.8, 0.9, 1,0, 1,1, 1,2, 1,3 0, 1, 2
ETM+ bands Weight for each band Distance weighting parameter t
t
df
Sig. (2-tailed)
−1.325 −0.058
624 0.186 624 0.953
Upper
0.459 0.602
−1.509 0.293 −1.147 1.218
0.619
−1.214
1.215
0.000 624 1.000
estimation. With the sampling intensity of the 625-field inventory, we tested every combination of k values (ktest = 2, 3…12) as well of distances ranging from 1 km to 100 km. The optimal number of the nearest neighbours' k was eight. Meanwhile, a geographical reference area with a radius of 55 km, ETM+ bands (1, 2, 3, 4, 5, 7) with their optimal weights, and inverse distance function (t = 1) were chosen when the RMSEcv (11.48%) and bias (1.26) were minimal. The predicted vegetation cover ranged from 11% to 100% (Fig. 2). The observed and predicted vegetation covers by the k-NN method were in the same group with a correlation coefficient of 0.704 (Pairedsample t-test, p b 0.001). The mean value difference between the observed and predicted variable was −0.608%. The predicted mean value was not significantly different from the observed one (t-test, p N 0.05). The predictive performance of the k-NN method was better than that of both a multivariate exponential regression model and the NDVI (Table 2). Of these three methods for predicting vegetation cover, the k-NN technique showed the least standard deviation and standard error. 4.2. Soil loss and its influencing factors The predicted soil loss has been classified according to the classification criteria of water erosion intensity (MWR, 1997) into the following six erosion intensity categories: negligible, slight, moderate, severe, very severe, and extremely severe. Table 3 shows the area and proportion of each soil loss class. Under the present conditions, about 75% of the land in the watershed was classified as stable, 10% was at the level of slight and moderated erosion, whereas 15% showed severe, very severe, and extremely severe erosion loss. Non-parametric correlation tests were used to verify the relationship between different factors and soil loss. The result showed that the LSfactor correlated positively with soil loss (correlation coefficient 0.735) (p b 0.01, Spearman's rho tests) as did the C-factor (0.605, p b 0.01) and the R-factor (0.079, p b 0.01). In contrast, soil erodibility factor K showed a negative correlation coefficient −0.047 (p b 0.01). The phenomenon revealed that it is not the soil K-factor that
Table 3 Derivation of the ordinal categories of soil erosion class and the area and proportion of each category Soil loss class
Intensity
Annual mean erosion modulus (t km− 2 yr− 1)
Area (km2)
Proportion (%)
1 2 3 4 5 6
Negligible Slight Moderate Severe Very severe Extremely severe
b500a (1000) 500a (1000)–2500 2500–5000 5000–10,000 10,000–15,000 N15,000 No data
5530 452 317 362 214 507 19
74.7 6.1 4.3 4.9 2.9 6.9 0.2
a This number is applied in mountainous areas, and the number in parentheses is applied in other areas. Water erosion intensities are typically divided into six classes according to standards set by the Ministry of Water Resources, PRC. (MWR, 1997).
P. Zhou et al. / Catena 75 (2008) 319–325
323
Fig. 3. Soil erosion class a. with current vegetation cover, b. a scenario with cover management factor (C-factor) equal to 1, c. a scenario with C-factor equal to 0.001.
contributes negatively to soil erosion, but rather that, at the landscape level, the value of soil erodibility showed a spatially different pattern of soil erosion. The low correlation coefficient between the R-factor
and soil loss was also probably due to their different patterns on a large scale. In general, the LS-factor and the C-factor are the most important factors affecting soil loss in this mountainous watershed.
Fig. 4. Different vegetation cover scenarios on areas of high erosion risk: a. high-risk area predicted under current vegetation cover, b–f. soil loss predicted for the high-risk areas, if restore the vegetation cover below 40% to 40%, below 50% to 50%, below 60% to 60%, cover below 70% to 70%, below 78.3% to 78.3%, respectively.
324
P. Zhou et al. / Catena 75 (2008) 319–325
4.3. Effect of C-factor on soil erosion loss Cover management factor C is the ratio which compares the soil loss under vegetation cover to that from bare soil. It is therefore a ratio from a value of zero, when the soil is completely protected, to a value of one for bare soil. Through a series of computations using map algebra, we generated maps to show what could result from different cover management factors. Under the present vegetation cover, around 15% of the area showed high erosion loss (5% severe, 3% high severe, and 7% extreme severe soil loss) (Fig. 3a). If all the land were without vegetation protection (C-factor = 1), then around 98% of the watershed would suffer extremely severe erosion loss (Fig. 3b). In such a case, mudslides and landslides could easily occur after a rainfall. If all the land were covered by dense plants (C-factor = 0.001), a sharp reduction in severe or higher erosion loss would result. In this case, only 0.4% of the watershed would exhibit severe or higher erosion loss (Fig. 3c). The scenarios showed that vegetation cover could efficiently protect soil against erosion loss. Good plant cover is capable of preventing soil erosion and reducing landslides as well. Removing vegetation can greatly increase soil erosion, particularly in mountainous areas. 4.4. Restoration scenario for areas at high erosion risk Fig. 4a shows the area at high-risk for erosion, which combined severe, very severe and extremely severe soil loss classes, in red. This area comprised 14.6% of the study area, with an area of 1083 km2 in the watershed. Most of the high-risk areas had significant less vegetation cover than did areas at lower risk for erosion. Fig. 4b–f shows different vegetation restoration scenarios. If high-risk areas with a vegetation cover of less than 40% were restored to 40%, only a slight reduction in higher levels of soil erosion resulted. Similar results were obtained when vegetation cover was restored to 50% or 60% (Fig. 4c and d). Restoring vegetation cover to 70% significantly reduced extremely severe erosion (Fig. 4e). Restoring the vegetation cover to at least 78.3% for all the land would yield a C-factor of at least 0.001, a sharp reduction in high erosion risk, and would result in negligible soil loss for around 41% of the area, slight soil loss for 56%, and high erosion risk for only 0.1% (around 8 km2) of the entire study area (Fig. 4f). The result suggested that in a mountainous watershed, high erosion tends to occur when more than 30% of the soil is exposed; a soil vegetation cover of more than 78% can greatly reduce erosion by water. 5. Discussion and conclusion The cover management factor has often been used to assess the effect of vegetation on soil erosion, and the C-factor can be ranked qualitatively based on vegetation types (Morgan, 1995), calculated from NDVI (De Jong, 1994; Thiam, 2003; Wu et al., 2004; Yu et al., 2006), or predicted by a regression function using Landsat TM or ETM+ bands (Hall et al., 2006). In this study, we employed the k-NN method for vegetation cover estimation and tested the performance of predictions by this method to those by NDVI and multivariate regression. Compared to other methods of estimating the C-factor, the k-NN method showed two main advantages: the method used an optimal combination of information derived from satellite images and ground measurements and simultaneously utilized the ground information for cross-validation. The accuracy of vegetation cover estimation is comparable to that of other similar studies for canopy cover or crown cover estimations. The RMSE of cross-validation for vegetation cover estimation in this study is 11.48% (ρ = 0.704). Hall et al. (2006) uses a multivariate exponential regression function to estimate crown closure using Landsat ETM+ bands 3, 4 and 7, which results in a RMSE of 12.0%. Among the factors affecting soil erosion in this mountainous watershed, the C-factor and LS-factor are the most important. Removal
of vegetation can greatly increase runoff and soil erosion, particularly in mountainous areas. The scenarios showed that almost all areas would be subject to extremely serious erosion loss without the protection of vegetation under the current landform. Greater vegetation cover can considerably reduce the loss of soil erosion. Only a small area of high-risk soil erosion will remain regardless of slope steepness and slope length, even if the vegetation cover exceeds 78.3%. Studies of the Loess Plateau show that vegetation is the key factor affecting soil erosion (Zheng, 2006). Under the same geographical landform, vegetation has a more noteworthy influence on soil erosion. The presence of a vegetation cover can increase infiltration, reduce surface runoff, and thus significantly retard sheet erosion (Woo and Luk, 1990). With reduced vegetation cover, however, runoff and soil erosion can greatly increase, resulting in flooding and mudslides (Varis and Vakkilainen, 2001; Sidle et al., 2004). Besides its role in carbon sequestration, vegetation plays an important role in protecting soil from erosion loss. Soil conservation programmes, such as the Grain for Green programme in China (Ye et al., 2003), have been established to reduce soil erosion. Using ArcGIS software, the RUSLE can be applied to predict potential soil loss quantitatively and spatially for each pixel. The use of GIS technology enables the quantitative spatial modelling of soil erosion by water, so that the areas at high erosion risk can be identified for the implementation of conservation measures. This study identified the high erosion risk areas where soil conservation practices are needed. The identification of areas at high-risk for erosion is of prime importance in soil conservation planning. Implementing soil conservation measures in high-risk areas not only increases the effectiveness of reducing soil loss, but also of reducing the cost of soil conservation. Various soil conservation scenarios can be evaluated easily through GIS dataset manipulation. Large-scale vegetation restoration or rehabilitation could offer a potential way to improve soil stability and reduce the soil loss by means of selecting suitable woody plant species for different soils at different elevations, thus taking advantage of existing forests (Zhou et al., 2007). Soil loss hazards may be alleviated if combined conservation measures, including terraces and contour tillage, are implemented in the watershed (Shi et al., 2004). Acknowledgements The authors thank the Academy of Finland for its financial support, the Chengdu Institute of Biology of the Chinese Academy of Sciences for making arrangements during the field work, and the staff of both the Sichuan Academy of Forestry and of the Maoxian Ecological Station for their kind assistance with the field survey. References De Jong, S.M., 1994. Derivation of vegetative variables from a Landsat TM image for modelling soil erosion. Earth Surface Processes and Landforms 19, 165–178. Food and Agriculture Organization of United Nations (FAO), 2006. Guidelines for Soil Description, Fourth edition. FAO, Rome. Franco-Lopez, H., Ek, A.R., Bauer, M.E., 2001. Estimation and mapping of forest stand density, volume, and cover type using the k-nearest neighbors method. Remote Sensing of Environment 77, 251–274. GLCF (Global Land Cover Facility) - www.landcover.org as a source of:- U.S. Geological Survey. 10th July 2002, Landsat ETM+, Scene, WRS-2 Path 130, Row 037, Orthorectified Geocover, Sioux Falls, South Dakota: USGS. - U.S. Geological Survey. 10th July 2002, Landsat ETM+, Scene, WRS-2 Path 130, Row 038, Orthorectified Geocover, Sioux Falls, South Dakota: USGS. Grønsten, H.A., Lundekvam, H., 2006. Prediction of surface runoff and soil loss in southeastern Norway using the WEPP Hillslope model. Soil and Tillage Research 85, 186–199. Hall, R.J., Skakun, R.S., Arsenault, E.J., Case, B.S., 2006. Modeling forest stand structure attributes using Landsat ETM+ data: application to mapping of aboveground biomass and stand volume. Forest Ecology and Management 225, 378–390. Janey, J.L., Block, W.M., 1994. A comparison of two techniques for measuring canopy closure. Western Journal of Applied Forestry 9, 21–23. Katila, M., Tomppo, E., 2001. Selecting estimation parameters for the Finnish multisource National Forest Inventory. Remote Sensing of Environment 76, 16–32.
P. Zhou et al. / Catena 75 (2008) 319–325 Langran, K.J., 1983. Potential for monitoring soil erosion features and soil erosion modelling components from remotely sensed data. Proceedings of IGARSS'83. IEEE, San Francisco, CA, pp. 2.1–2.4. Millington, A.C., Townshend, J.R.G., 1984. Remote sensing applications in African erosion and sedimentation studies. In: Walling, D.E., Foster, S.S.D., Wurzel, P. (Eds.), Challenges in African Hydrology and Water Resources: Proceedings of the Harare Symposium. . IAHS Publication, vol. 144. IAHS Press, pp. 373–384. Morgan, P.R.C., 1995. Soil Erosion and Conservation. Longman Group Limited, Essex, UK. MWR (Ministry of Water Resources, PRC), 1997. Standards for Classification and Gradation of Soil Erosion (SL190-96). China Hydraulic and Hydropower Press, Beijing. MWR (Ministry of Water Resources, PRC), 1999. National Soil and Water Conservation and Ecological Environment Construction Plan 1998–2050 (in Chinese). Chinese Water Conservation and Hydropower Press, Beijing. Renard, K.G., Foster, G.R., Weesies, G.A., McCool, D.K., Yoder, D.C., 1997. Predicting Soil Erosion by Water: A Guide to Conservation Planning with the Revised Soil Loss Equation (RUSLE). U.S. Dept. of Agriculture. Agric. Handbook No. 703, 404pp. Shi, Z.H., Cai, C.F., Ding, S.W., Wang, T.W., Chow, T.L., 2004. Soil conservation planning at the small watershed level using RUSLE with GIS: a case study in the Three Gorge Area of China. Catena 55, 33–48. Sidle, R.C., Taylor, D., Lu, X.X., Adger, W.N., Lowe, D.J., de Lange, W.P., Newnham, R.M., Dodson, J.R., 2004. Interactions of natural hazards and society in Austral–Asia: evidence in past and recent records. Quaternary International 118–119, 181–203. Tarboton, D.G., 1997. A new method for the determination of flow directions and contributing areas in grid digital elevation models. Water Resources Research 33, 309–319. Thiam, A.K., 2003. The causes and spatial pattern of land degradation risk in southern Mauritania using multitemporal AVHRR-NDVI imagery and field data. Land Degradation & Development 14, 133–142. Tokola, T., 2000. The influence of field sample data location on growing stock volume estimation in Landsat TM-based forest inventory in eastern Finland. Remote Sensing of Environment 74, 422–431. Tomppo, E., 1991. Satellite image-based national forest inventory of Finland. Proceedings of the Symposium on Global and Environmental Monitoring, Techniques and Impacts, 17–21 Sep. 1990. Victoria, British Columbia, Canada, International Archives of Photogrammetry and Remote Sensing 28, pp. 419–421. Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sensing of Environment 8, 127–150. Varis, O., Vakkilainen, P., 2001. China's 8 challenges to water resources management in the first quarter of the 21st Century. Geomorphology 41, 93–104.
325
Warren, S.D., Diersing, V.E., Thompson, P.J., Goran, W.D., 1989. An erosion-based land classification system for military installations. Environmental Management 13, 251–257. Wischmeier, W.H., Smith, D.D., 1978. Predicting Rainfall Erosion Losses — A Guide to Conservation Planning. U.S. Dept. of Agriculture. Agric. Handbook No. 537, 58 pp. Woo, M., Luk, S., 1990. Vegetation effects on soil and water losses on weathered granitic hillslopes, south China. Physical Geography 11, 1–16. Wu, B., Zhou, Y., Huang, J., Tian, Y., Huang, W., 2004. Spatial Pattern of Soil and Water Loss and Its Affecting Factors Analysis in the Upper Basin of Miyun Reservoir. IEEE, pp. 4700–4702. Wu, Y., Su, Z.X., Fang, J.Y., 2003. Study on causes and ecological renewal of arid and warm valley of Upper Minjiang River. Journal of China West Normal University 24, 276–281. Yang, Y.S., Shi, D.M., 1994. Study on Soil Erosion in the Three Gorge Area of the Changjiang River. Southeast University Press, Nanjing. (In Chinese). Ye, Y.Q., Chen, G.J., Fan, H., 2003. Impacts of the “Grain for Green” project on rural communities in the Upper Min River Basin, Sichuan, China. Mountain Research and Development 23, 345–352. Yu, X., Zhang, X., Li, J., Zhang, M., Xie, Y., 2006. Effect of vegetation cover and precipitation on the process of sediment produced by erosion in a small watershed of loess region. Acta Ecologica Sinica 26, 1–8. Zhang, X., Zhang, Y., Wen, A., Feng, M., 2003. Assessment of soil losses on cultivated land by using the 137Cs technique in the Upper Yangtze River Basin of China. Soil and Tillage Research 69, 99–106. Zheng, F., 2006. Effect of vegetation changes on soil erosion on the Loess Plateau. Pedosphere 16, 420–427. Zhou, P., Nieminen, J., Tokola, T., Luukkanen, O., Oliver, T., 2006. Large scale soil erosion modeling for a mountainous watershed. In: Martin-Duque, J.F., Brebbia, C.A., Emmanouloudis, D.E., Mander, E. (Eds.), Geo-Environment & Landscape Evolution II. WIT press, UK, pp. 55–67. Zhou, P., Luukkanen, O., Tokola, T., Nieminen, J., 2007. Vegetation Dynamics and Forest Landscape Restoration in the Upper Min River Watershed, Sichuan, China. Restoration Ecology (OnlineEarly Articles). doi:10.1111/j.1526-100X.2007.00307.x. Zhou, P., 2008. Landscape-scale Soil Erosion Modelling and Ecological Restoration for a Mountainous Watershed in Sichuan, China. Viikki Tropical Resources Institute, Helsinki.
Contents lists available at ScienceDirect
Catena j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / c a t e n a
Effect of vegetation cover on soil erosion in a mountainous watershed P. Zhou a,⁎, O. Luukkanen b, T. Tokola c,d, J. Nieminen c a
College of Resources and Environment, Northwest Agriculture and Forestry University. P. O. Box 81, 712100, Shaanxi, China Viikki Tropical Resources Institute, Department of Forest Ecology, University of Helsinki, P. O. Box 27, FIN-00014, Helsinki, Finland c Department of Forest Resource Management, University of Helsinki, P. O. Box 27, FIN-00014, Helsinki, Finland d Faculty of Forestry, University of Joensuu, P. O. Box 111, FIN-80101, Joensuu, Finland b
a r t i c l e
i n f o
Article history: Received 3 February 2008 Received in revised form 17 June 2008 Accepted 28 July 2008 Keywords: k-NN Restoration RUSLE Soil erosion Upper Min River watershed Vegetation cover
a b s t r a c t We applied the Revised Soil Loss Equation (RUSLE) to assess levels of soil loss in a Geographic Information System (GIS). In this study, we used the k-NN technique to estimate vegetation cover by integrating Landsat ETM+ scenes and field data with optimal parameters. We evaluated the root mean square errors and significance of biases at the pixel level in order to determine the optimal parameters. The accuracy of vegetation cover estimation by the k-NN technique was compared to that predicted by a regression function using Landsat ETM+ bands and field measurements as well as to that predicted by the Normalized Difference Vegetation Index (NDVI). We used a regression equation to calculate the cover management (C) factor of the RUSLE from vegetation cover data. On the basis of the quantitative model of soil erosion, we explored the relationship between soil loss and its influencing factors, and identified areas at high erosion risk. The results showed that the k-NN method can predict vegetation cover more accurately for image pixels at the landscape level than can the other two methods examined in this study. Of those factors, the C-factor is one of the most important affecting soil erosion in the region. Scenarios with different vegetation cover on high-risk areas showed that greater vegetation cover can considerably reduce the loss of soil erosion. The k-NN technique provides a new method to estimate the C-factor for RUSLE erosion mapping. The quantitative model of different vegetation cover scenarios provides information on how vegetation restoration could reduce erosion. © 2008 Elsevier B.V. All rights reserved.
1. Introduction Soil erosion by water is a worldwide environmental problem which degrades soil productivity and water quality, causes sedimentation and increases the probability of flood. The Upper Min River (UMR) watershed in the Upper Yangtze Basin is an environmentally fragile area due to deforestation and soil erosion. Up to 44% of the land has been described as degraded in the area (Ye et al., 2003; Wu et al., 2003). Different methods have been developed to detect eroded areas and to assess erosion loss. For instance, the qualitative classification of soil erosion can be assessed by FAO (2006), and large and middle-size eroded areas can be directly identified with Landsat and SPOT imagery (Langran, 1983; Millington and Townshend, 1984); the quantitative soil loss can be modeled with the universal soil loss equation (USLE) (Wischmeier and Smith, 1978), its revised version (RUSLE) (Renard et al., 1997), the Soil Erosion Model for Mediterranean regions (SEMMED) (De Jong,1994), 137CS techniques (Zhang et al., 2003), and the Water Erosion Prediction Project (WEPP) hill slope model (Grønsten and Lundekvam, 2006). Of these models, the USLE and RUSLE are the most widely used, providing a convenient tool for soil loss evaluation by taking into consideration rainfall, ⁎ Corresponding author. Tel.: +86 29 87080321; fax: +86 29 87080055. E-mail address: [email protected] (P. Zhou). 0341-8162/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.catena.2008.07.010
topography, conservation support practice, soil, and vegetation. Although they are empirical models for assessing long-term averages of sheet and rill erosion based on plot data, the calculation of their factors can be improved and adapted to enable their application to various spatial scales and region sizes in different environments using GIS (Warren et al., 1989). For better prediction of soil loss in complex terrain, the modelling of dispersed flow over the surface can improve the slope length and steepness factor (LS) (Tarboton, 1997). Of these six RUSLE factors, the cover management factor (C-factor) is an important one affecting soil erosion in a given region. However, modelling the C-factor for complex terrain is difficult. The C-factor in the soil loss equation is defined as the ratio of soil loss from land cropped under specified conditions to the corresponding loss from clean-tilled, continuous fallow (Wischmeier and Smith, 1978). Land use classification is often used to map vegetation types that differ in their effectiveness to protect the soil. After classification, C-factors are assigned according to a qualitative ranking of vegetation types (Wischmeier and Smith, 1978; Morgan, 1995). However, the proportional vegetation cover varied widely even for the same vegetation type. The direct application of the C-factor from the RUSLE is based on prior land use (PLU), canopy cover (CC), surface cover (SC), surface roughness (SR), and soil moisture (SM) (Renard et al., 1997). However, assessing PLU, CC, SC, SR, and SM simultaneously for a large area not covered mainly by agricultural lands is difficult. The Normalized Difference Vegetation Index (NDVI), defined
320
P. Zhou et al. / Catena 75 (2008) 319–325
as the near infrared reflection minus the red reflection divided by the sum of the two (Tucker, 1979), was calculated from satellite images for C-factor estimation (Thiam, 2003; Wu et al., 2004) because green vegetation is more reflective in the near infrared part and less reflective in the red part of the spectrum. However, De Jong (1994) found that the relationship between Landsat-derived spectral indices and vegetation attributes was quite poor for Mediterranean France. The non-parametric k-nearest neighbour (k-NN) technique, widely used in a variety of forest estimation and biomass mapping applications over the years (Tokola, 2000; Franco-Lopez et al., 2001; Katila and Tomppo, 2001), provides a new method to map proportional vegetation cover. The k-NN technique is a non-parametric approach to predicting values of point variables on the basis of similarity in a covariate space between one point and other points with observed values for the variables (Tomppo, 1991). Using the k-NN method, we cannot only predict variables such as vegetation cover, aboveground biomass, and stand volume or parameters for each pixel, but also examine the covariance structure of field variables and produce pixel-by-pixel maps. In this study, we used the k-NN method to estimate vegetation cover data and produced a pixel-by-pixel map by integrating the satellite images and field data with optimal parameters. The major objectives of this study were to produce a vegetation cover map using the k-NN method and testing the performance of this method, to quantitatively evaluate soil erosion at 25-m resolution for a 7400 km2 area of the watershed using the RUSLE model with predicted factors, especially an improved C-factor, to test the relationship between soil erosion and its influencing factors, as well as to model soil loss risk under different vegetation cover scenarios.
2. Study area The study area is located in the Upper Min River (UMR) watershed (Fig. 1), with an area of 23,040 km2. The target watershed is in Sichuan province, Southwest China, between 31°–33° N and 102°–104° E. The Min River is also one of the most important tributaries of the Upper Yangtze River. The climate is governed by the northeast and southwest monsoons. A complex topography, with elevation ranges from 600 m to more than 6000 m a.s.l. (above sea level), results in steep gradients of rainfall on both spatial and temporal scale. The annual precipitation ranges from 405 mm to 1950 mm in different parts of the watershed, and around half of the precipitation falls in July, August and September (Zhou, 2008). The upper reaches of the Yangtze River are the region suffering from serious water erosion in China. Around 1.56 billion tons of soil is eroded in this region each year (MWR, 1999). The UMR watershed is fragile to water erosion due to steep terrain and deforestation. The forest cover had declined from 50% to 30% by 1950 and to 18.8% by the 1980s, and to only 5%–7% along the main river (Wu et al., 2003). At present, around 21% of the entire watershed has forest cover. The vegetation here ranged from subtropical evergreen broadleaved forest to alpine meadows. Our 625 inventory plots were randomly placed in the middle and upper reaches of the UMR watershed (Fig. 1) over an area of about 7400 km2. The vegetation ranges from subtropical evergreen broadleaved forest to alpine meadows. Slope angles range from 0 to 77.2° with a mean of 25.9° (Zhou et al., 2006). The area comprises three main soil orders (Alfisol, Semi-alfisol, and Semi-aquatic) and faces
Fig. 1. Location of the research area and the sample plots.
P. Zhou et al. / Catena 75 (2008) 319–325
many environmental problems such as land and soil degradation, soil erosion, mud- and landslides, earthquakes, and the growing desertification of some valleys along the main Min River. 3. Materials and methods 3.1. Soil erosion by RUSLE The RUSLE has served to evaluate and predict soil erosion in the region on a pixel-by-pixel basis. The soil loss (A) due to water erosion per unit area per year (Mg ha yr− 1) was quantified using the RUSLE with the following equation: A ¼ R K LS C P
ð1Þ
where A is the average soil loss due to water erosion, R the rainfall and runoff erosivity factor (MJ mm ha− 1 h− 1 yr− 1), K the soil erodibility factor (Mg h MJ− 1 mm− 1), L the slope length factor (contributing area), S the slope steepness factor, C the cover and management practice factor, and P the support practice. We generated a digital elevation model (DEM) with a 25-m cell size grid using both contour lines and stream data to ensure hydrological accuracy. The data used to generate R, LS, and K included precipitation data, DEM, satellite images, inventory data, digitised stream data, soil map and soil erodibility data. Zhou et al. (2006) have separately described the method of raster maps generation with regard to R, LS, and K factors. Precipitation data measured with rain gauge from 1998 to 2002 was collected from 40 meteorological stations within and surrounding the research area. A multivariate geostatistic cokriging model was used to produce the raster map of R-factor. We used the standard for soil erosion intensity issued by the Ministry of Water Resources, China (MWR, 1997) to classify soil loss in this specific watershed. The standard was determined mainly based on field experiments in China and was guaranteed to be applicable for the whole China. According to the predicted soil loss amount, the potential soil loss was divided into six ordinal classes. We then calculated the area and percentage of each soil loss class by generating zonal statistics in ArcGIS. 3.2. C-factor estimation We employed the k-NN technique to predict vegetation cover for image pixels and used consecutive Landsat ETM+ scenes WRS2 130/ 037 and 130/038 from 10 July 2002 (GLCF, 2002) to estimate vegetation cover. To ensure compatibility between the images and the ground data, each image was rectified and georeferenced to the Universal Transverse Mercator system (UTM): WGS_1984_UTM_Zone_48N. The vegetation cover image was constructed using ETM+ bands 1, 2, 3, 4, 5 and 7. We recorded vegetation covers at three levels (canopy cover, under canopy cover, and overall vegetation cover) from 625 sample plots and measured the canopy cover with a spherical densiometre (Janey and Block, 1994). The field inventory was conducted in June, July and August 2004 using strata-delineated sampling with the aid of the global positioning system (GPS). Besides the measurement on vegetation cover at three levels, other recordings were also made such as the elevation, slope, aspect, forest type, soil type, plot condition, land use. More than 5000 field pictures were taken from the field and later on the pictures have been also used for visually checking the results. We calculated the C-factor of the RUSLE from the predicted vegetation cover image using a regression function Eq. (2). The equation was build by using over 200 soil loss ratios measured from 30 runofferosion plots under both natural and simulated rainfall events in the Three Gorges area of the Yangtze River basin (Yang and Shi, 1994). C ¼ 0:6508−0:343 log c 0 < c < 78:3%
ð2Þ
321
3.3. The k-NN technique for vegetation cover estimation The data for the k-NN method used in this study included a georeferenced satellite image and field plot data containing the following values: the x and y coordinate locations of plots, their corresponding satellite image spectral values from six bands, and total vegetation cover. To predict the vegetation cover using k-NN technique, consider an ETM+ pixel to be a point, let Y denote vegetation cover. The estimation of variable Y for pixel j can be computed from Eq. (3). Yˆ j ¼
! !−1 k k 1 ∑ wji ∑ wji Yji k i¼1 i¼1
ð3Þ
where k is a predetermined constant, 1 ≤ k ≤ n, {wji} are the weights of the points to be selected, and Yji = 1 ,…, k, is the value of variable Y in the sample plot i corresponding to the pixel j, which is the ith closest pixel in the spectral space to the pixel j. The geographical distance ranging from 1 km to 100 km to the potential nearest neighbours has been tested. With the leave-one-out method, a k-NN prediction (Yˆi) is sequentially obtained for each Yi,i = 1 ,… k, while Yi itself are not included in the mean forming of its own k-NN prediction. The quality of the predictions can be evaluated by comparing the observations Yji = 1 ,…, k and the corresponding predictions Yˆj. Specific parameters for the k-NN technique must be estimated before implementation. The particular spectral bands should be selected to calculate distances dji between spectral band Xj and each element of the set (Xi,i = 1,… ,n). Distances between neighbours were computed using weighted Euclidean distance metrics Eq. (4). dji ¼
M 2 ∑ vm Xjm −Xim
1=2
m¼1
ð4Þ
where {vm} are variable weights for the feature m, and M are the selected bands. Variable weights are used in the equation because not all the features in the feature space share the same degree of influence on the prediction of vegetation cover for a given pixel. Finally {wji} must be tested. Three different weighting functions such as equal weights , inverse distance weighting, and inverse to the square of the distance weighting were used to compute {wji} .These weights are computed by choosing t = 0, 1, or 2 respectively Eq. (5). wji ¼ dji−t
ð5Þ
The accuracy of the estimate on vegetation cover can be examined using the root mean square error of cross-validation (RMSECV): " RMSECV ¼
1 n ∑ n i¼1
2 #1=2 −i ˆ Yi − Y i
ð6Þ
where Yˆi− i is the predicted value of the ith observation using the leaveone-out method without considering observation i. The optimal number of k as well as of the number of bands and weights were chosen when RMSE and bias were minimal. We compared the predictive performance of the k-NN method to that of a multivariate exponential regression model, as well as to that of the NDVI. The method explained by Hall et al. (2006), which tests have shown to yield the best model performance for predicting height and crown closure, was used to construct the multivariate model (Y = exp (B0 + B1X1 + … + BnXn)). Data used for building the multivariate exponential regression model in this study were satellite image spectral values from six bands, and field recordings on vegetation cover with GPS coordinates. The NDVI was computed from spectral data from red band 3 (RED) (wavelength 0.63–0.69 µm) and near infrared band 4 (NIR) (wavelength 0.75–0.90 µm) from the Landsat ETM+ (NDVI = (NIR − RED)/(NIR + RED)) (Tucker, 1979). A regression
322
P. Zhou et al. / Catena 75 (2008) 319–325 Table 2 Paired-sample t-test on observed and predicted vegetation cover using k-NN, multivariate regression and NDVI Methods
Paired differences (%) Mean
Std. Std. error 95% deviation confidence interval Lower
k-NN −0.608 11.472 Multivariate 0.035 15.054 regression NDVI 0.0002 15.465
Fig. 2. Vegetation cover map at 25 × 25 m pixel size predicted with k-NN method.
function was then built between calculated NDVI and measured vegetation cover for those sample plots. The paired-sample t-test (SPSS 12.0.1) was used to check the accuracy of the estimate by comparing the observed and predicted vegetation cover with these three methods. 3.4. Scenarios of vegetation cover on soil erosion Through a series of computations using map algebra based on Eqs. (1) and (2), we generated maps to show what effects could result from different cover management factors using ArcGIS software. First of all, soil loss intensities were compared under the current vegetation cover as well as under scenarios of no vegetation protection (C = 1) and good vegetation protection (C = 0.001). Five different vegetation cover restoration scenarios, such as a restoration of vegetation cover of less than 40%, 50%, 60%, 70%, or 78.3% for each respective value, were simulated to calculate how different vegetation cover affected the areas at high-risk for soil loss. Zonal statistics were used to calculate the proportions of different soil erosion intensities under each scenario. 4. Results 4.1. Vegetation cover map Vegetation cover with a 25 × 25-m pixel size was produced using the non-parametric k-nearest neighbour (k-NN) multi-source estimation method by integrating the satellite images and field data with optimal parameters (Fig. 2). The set of parameters chosen for the k-NN method to predict the vegetation cover included the image bands, the weight for each band, the distance, the number of nearest neighbours, the value of k, and the geographical reference area of the nearest field plots (Table 1). The leave-one-out cross-validation method was applied to calculate the RMSE and the average biases of predictions at the single-pixel level for different combinations of the k-NN
Table 1 Tested parameters for vegetation cover estimation using k-NN method Parameters
Values
K Geographical distance (km)
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 1, 5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95, 100 1, 2, 3, 4, 5, 7 0.7, 0.8, 0.9, 1,0, 1,1, 1,2, 1,3 0, 1, 2
ETM+ bands Weight for each band Distance weighting parameter t
t
df
Sig. (2-tailed)
−1.325 −0.058
624 0.186 624 0.953
Upper
0.459 0.602
−1.509 0.293 −1.147 1.218
0.619
−1.214
1.215
0.000 624 1.000
estimation. With the sampling intensity of the 625-field inventory, we tested every combination of k values (ktest = 2, 3…12) as well of distances ranging from 1 km to 100 km. The optimal number of the nearest neighbours' k was eight. Meanwhile, a geographical reference area with a radius of 55 km, ETM+ bands (1, 2, 3, 4, 5, 7) with their optimal weights, and inverse distance function (t = 1) were chosen when the RMSEcv (11.48%) and bias (1.26) were minimal. The predicted vegetation cover ranged from 11% to 100% (Fig. 2). The observed and predicted vegetation covers by the k-NN method were in the same group with a correlation coefficient of 0.704 (Pairedsample t-test, p b 0.001). The mean value difference between the observed and predicted variable was −0.608%. The predicted mean value was not significantly different from the observed one (t-test, p N 0.05). The predictive performance of the k-NN method was better than that of both a multivariate exponential regression model and the NDVI (Table 2). Of these three methods for predicting vegetation cover, the k-NN technique showed the least standard deviation and standard error. 4.2. Soil loss and its influencing factors The predicted soil loss has been classified according to the classification criteria of water erosion intensity (MWR, 1997) into the following six erosion intensity categories: negligible, slight, moderate, severe, very severe, and extremely severe. Table 3 shows the area and proportion of each soil loss class. Under the present conditions, about 75% of the land in the watershed was classified as stable, 10% was at the level of slight and moderated erosion, whereas 15% showed severe, very severe, and extremely severe erosion loss. Non-parametric correlation tests were used to verify the relationship between different factors and soil loss. The result showed that the LSfactor correlated positively with soil loss (correlation coefficient 0.735) (p b 0.01, Spearman's rho tests) as did the C-factor (0.605, p b 0.01) and the R-factor (0.079, p b 0.01). In contrast, soil erodibility factor K showed a negative correlation coefficient −0.047 (p b 0.01). The phenomenon revealed that it is not the soil K-factor that
Table 3 Derivation of the ordinal categories of soil erosion class and the area and proportion of each category Soil loss class
Intensity
Annual mean erosion modulus (t km− 2 yr− 1)
Area (km2)
Proportion (%)
1 2 3 4 5 6
Negligible Slight Moderate Severe Very severe Extremely severe
b500a (1000) 500a (1000)–2500 2500–5000 5000–10,000 10,000–15,000 N15,000 No data
5530 452 317 362 214 507 19
74.7 6.1 4.3 4.9 2.9 6.9 0.2
a This number is applied in mountainous areas, and the number in parentheses is applied in other areas. Water erosion intensities are typically divided into six classes according to standards set by the Ministry of Water Resources, PRC. (MWR, 1997).
P. Zhou et al. / Catena 75 (2008) 319–325
323
Fig. 3. Soil erosion class a. with current vegetation cover, b. a scenario with cover management factor (C-factor) equal to 1, c. a scenario with C-factor equal to 0.001.
contributes negatively to soil erosion, but rather that, at the landscape level, the value of soil erodibility showed a spatially different pattern of soil erosion. The low correlation coefficient between the R-factor
and soil loss was also probably due to their different patterns on a large scale. In general, the LS-factor and the C-factor are the most important factors affecting soil loss in this mountainous watershed.
Fig. 4. Different vegetation cover scenarios on areas of high erosion risk: a. high-risk area predicted under current vegetation cover, b–f. soil loss predicted for the high-risk areas, if restore the vegetation cover below 40% to 40%, below 50% to 50%, below 60% to 60%, cover below 70% to 70%, below 78.3% to 78.3%, respectively.
324
P. Zhou et al. / Catena 75 (2008) 319–325
4.3. Effect of C-factor on soil erosion loss Cover management factor C is the ratio which compares the soil loss under vegetation cover to that from bare soil. It is therefore a ratio from a value of zero, when the soil is completely protected, to a value of one for bare soil. Through a series of computations using map algebra, we generated maps to show what could result from different cover management factors. Under the present vegetation cover, around 15% of the area showed high erosion loss (5% severe, 3% high severe, and 7% extreme severe soil loss) (Fig. 3a). If all the land were without vegetation protection (C-factor = 1), then around 98% of the watershed would suffer extremely severe erosion loss (Fig. 3b). In such a case, mudslides and landslides could easily occur after a rainfall. If all the land were covered by dense plants (C-factor = 0.001), a sharp reduction in severe or higher erosion loss would result. In this case, only 0.4% of the watershed would exhibit severe or higher erosion loss (Fig. 3c). The scenarios showed that vegetation cover could efficiently protect soil against erosion loss. Good plant cover is capable of preventing soil erosion and reducing landslides as well. Removing vegetation can greatly increase soil erosion, particularly in mountainous areas. 4.4. Restoration scenario for areas at high erosion risk Fig. 4a shows the area at high-risk for erosion, which combined severe, very severe and extremely severe soil loss classes, in red. This area comprised 14.6% of the study area, with an area of 1083 km2 in the watershed. Most of the high-risk areas had significant less vegetation cover than did areas at lower risk for erosion. Fig. 4b–f shows different vegetation restoration scenarios. If high-risk areas with a vegetation cover of less than 40% were restored to 40%, only a slight reduction in higher levels of soil erosion resulted. Similar results were obtained when vegetation cover was restored to 50% or 60% (Fig. 4c and d). Restoring vegetation cover to 70% significantly reduced extremely severe erosion (Fig. 4e). Restoring the vegetation cover to at least 78.3% for all the land would yield a C-factor of at least 0.001, a sharp reduction in high erosion risk, and would result in negligible soil loss for around 41% of the area, slight soil loss for 56%, and high erosion risk for only 0.1% (around 8 km2) of the entire study area (Fig. 4f). The result suggested that in a mountainous watershed, high erosion tends to occur when more than 30% of the soil is exposed; a soil vegetation cover of more than 78% can greatly reduce erosion by water. 5. Discussion and conclusion The cover management factor has often been used to assess the effect of vegetation on soil erosion, and the C-factor can be ranked qualitatively based on vegetation types (Morgan, 1995), calculated from NDVI (De Jong, 1994; Thiam, 2003; Wu et al., 2004; Yu et al., 2006), or predicted by a regression function using Landsat TM or ETM+ bands (Hall et al., 2006). In this study, we employed the k-NN method for vegetation cover estimation and tested the performance of predictions by this method to those by NDVI and multivariate regression. Compared to other methods of estimating the C-factor, the k-NN method showed two main advantages: the method used an optimal combination of information derived from satellite images and ground measurements and simultaneously utilized the ground information for cross-validation. The accuracy of vegetation cover estimation is comparable to that of other similar studies for canopy cover or crown cover estimations. The RMSE of cross-validation for vegetation cover estimation in this study is 11.48% (ρ = 0.704). Hall et al. (2006) uses a multivariate exponential regression function to estimate crown closure using Landsat ETM+ bands 3, 4 and 7, which results in a RMSE of 12.0%. Among the factors affecting soil erosion in this mountainous watershed, the C-factor and LS-factor are the most important. Removal
of vegetation can greatly increase runoff and soil erosion, particularly in mountainous areas. The scenarios showed that almost all areas would be subject to extremely serious erosion loss without the protection of vegetation under the current landform. Greater vegetation cover can considerably reduce the loss of soil erosion. Only a small area of high-risk soil erosion will remain regardless of slope steepness and slope length, even if the vegetation cover exceeds 78.3%. Studies of the Loess Plateau show that vegetation is the key factor affecting soil erosion (Zheng, 2006). Under the same geographical landform, vegetation has a more noteworthy influence on soil erosion. The presence of a vegetation cover can increase infiltration, reduce surface runoff, and thus significantly retard sheet erosion (Woo and Luk, 1990). With reduced vegetation cover, however, runoff and soil erosion can greatly increase, resulting in flooding and mudslides (Varis and Vakkilainen, 2001; Sidle et al., 2004). Besides its role in carbon sequestration, vegetation plays an important role in protecting soil from erosion loss. Soil conservation programmes, such as the Grain for Green programme in China (Ye et al., 2003), have been established to reduce soil erosion. Using ArcGIS software, the RUSLE can be applied to predict potential soil loss quantitatively and spatially for each pixel. The use of GIS technology enables the quantitative spatial modelling of soil erosion by water, so that the areas at high erosion risk can be identified for the implementation of conservation measures. This study identified the high erosion risk areas where soil conservation practices are needed. The identification of areas at high-risk for erosion is of prime importance in soil conservation planning. Implementing soil conservation measures in high-risk areas not only increases the effectiveness of reducing soil loss, but also of reducing the cost of soil conservation. Various soil conservation scenarios can be evaluated easily through GIS dataset manipulation. Large-scale vegetation restoration or rehabilitation could offer a potential way to improve soil stability and reduce the soil loss by means of selecting suitable woody plant species for different soils at different elevations, thus taking advantage of existing forests (Zhou et al., 2007). Soil loss hazards may be alleviated if combined conservation measures, including terraces and contour tillage, are implemented in the watershed (Shi et al., 2004). Acknowledgements The authors thank the Academy of Finland for its financial support, the Chengdu Institute of Biology of the Chinese Academy of Sciences for making arrangements during the field work, and the staff of both the Sichuan Academy of Forestry and of the Maoxian Ecological Station for their kind assistance with the field survey. References De Jong, S.M., 1994. Derivation of vegetative variables from a Landsat TM image for modelling soil erosion. Earth Surface Processes and Landforms 19, 165–178. Food and Agriculture Organization of United Nations (FAO), 2006. Guidelines for Soil Description, Fourth edition. FAO, Rome. Franco-Lopez, H., Ek, A.R., Bauer, M.E., 2001. Estimation and mapping of forest stand density, volume, and cover type using the k-nearest neighbors method. Remote Sensing of Environment 77, 251–274. GLCF (Global Land Cover Facility) - www.landcover.org as a source of:- U.S. Geological Survey. 10th July 2002, Landsat ETM+, Scene, WRS-2 Path 130, Row 037, Orthorectified Geocover, Sioux Falls, South Dakota: USGS. - U.S. Geological Survey. 10th July 2002, Landsat ETM+, Scene, WRS-2 Path 130, Row 038, Orthorectified Geocover, Sioux Falls, South Dakota: USGS. Grønsten, H.A., Lundekvam, H., 2006. Prediction of surface runoff and soil loss in southeastern Norway using the WEPP Hillslope model. Soil and Tillage Research 85, 186–199. Hall, R.J., Skakun, R.S., Arsenault, E.J., Case, B.S., 2006. Modeling forest stand structure attributes using Landsat ETM+ data: application to mapping of aboveground biomass and stand volume. Forest Ecology and Management 225, 378–390. Janey, J.L., Block, W.M., 1994. A comparison of two techniques for measuring canopy closure. Western Journal of Applied Forestry 9, 21–23. Katila, M., Tomppo, E., 2001. Selecting estimation parameters for the Finnish multisource National Forest Inventory. Remote Sensing of Environment 76, 16–32.
P. Zhou et al. / Catena 75 (2008) 319–325 Langran, K.J., 1983. Potential for monitoring soil erosion features and soil erosion modelling components from remotely sensed data. Proceedings of IGARSS'83. IEEE, San Francisco, CA, pp. 2.1–2.4. Millington, A.C., Townshend, J.R.G., 1984. Remote sensing applications in African erosion and sedimentation studies. In: Walling, D.E., Foster, S.S.D., Wurzel, P. (Eds.), Challenges in African Hydrology and Water Resources: Proceedings of the Harare Symposium. . IAHS Publication, vol. 144. IAHS Press, pp. 373–384. Morgan, P.R.C., 1995. Soil Erosion and Conservation. Longman Group Limited, Essex, UK. MWR (Ministry of Water Resources, PRC), 1997. Standards for Classification and Gradation of Soil Erosion (SL190-96). China Hydraulic and Hydropower Press, Beijing. MWR (Ministry of Water Resources, PRC), 1999. National Soil and Water Conservation and Ecological Environment Construction Plan 1998–2050 (in Chinese). Chinese Water Conservation and Hydropower Press, Beijing. Renard, K.G., Foster, G.R., Weesies, G.A., McCool, D.K., Yoder, D.C., 1997. Predicting Soil Erosion by Water: A Guide to Conservation Planning with the Revised Soil Loss Equation (RUSLE). U.S. Dept. of Agriculture. Agric. Handbook No. 703, 404pp. Shi, Z.H., Cai, C.F., Ding, S.W., Wang, T.W., Chow, T.L., 2004. Soil conservation planning at the small watershed level using RUSLE with GIS: a case study in the Three Gorge Area of China. Catena 55, 33–48. Sidle, R.C., Taylor, D., Lu, X.X., Adger, W.N., Lowe, D.J., de Lange, W.P., Newnham, R.M., Dodson, J.R., 2004. Interactions of natural hazards and society in Austral–Asia: evidence in past and recent records. Quaternary International 118–119, 181–203. Tarboton, D.G., 1997. A new method for the determination of flow directions and contributing areas in grid digital elevation models. Water Resources Research 33, 309–319. Thiam, A.K., 2003. The causes and spatial pattern of land degradation risk in southern Mauritania using multitemporal AVHRR-NDVI imagery and field data. Land Degradation & Development 14, 133–142. Tokola, T., 2000. The influence of field sample data location on growing stock volume estimation in Landsat TM-based forest inventory in eastern Finland. Remote Sensing of Environment 74, 422–431. Tomppo, E., 1991. Satellite image-based national forest inventory of Finland. Proceedings of the Symposium on Global and Environmental Monitoring, Techniques and Impacts, 17–21 Sep. 1990. Victoria, British Columbia, Canada, International Archives of Photogrammetry and Remote Sensing 28, pp. 419–421. Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sensing of Environment 8, 127–150. Varis, O., Vakkilainen, P., 2001. China's 8 challenges to water resources management in the first quarter of the 21st Century. Geomorphology 41, 93–104.
325
Warren, S.D., Diersing, V.E., Thompson, P.J., Goran, W.D., 1989. An erosion-based land classification system for military installations. Environmental Management 13, 251–257. Wischmeier, W.H., Smith, D.D., 1978. Predicting Rainfall Erosion Losses — A Guide to Conservation Planning. U.S. Dept. of Agriculture. Agric. Handbook No. 537, 58 pp. Woo, M., Luk, S., 1990. Vegetation effects on soil and water losses on weathered granitic hillslopes, south China. Physical Geography 11, 1–16. Wu, B., Zhou, Y., Huang, J., Tian, Y., Huang, W., 2004. Spatial Pattern of Soil and Water Loss and Its Affecting Factors Analysis in the Upper Basin of Miyun Reservoir. IEEE, pp. 4700–4702. Wu, Y., Su, Z.X., Fang, J.Y., 2003. Study on causes and ecological renewal of arid and warm valley of Upper Minjiang River. Journal of China West Normal University 24, 276–281. Yang, Y.S., Shi, D.M., 1994. Study on Soil Erosion in the Three Gorge Area of the Changjiang River. Southeast University Press, Nanjing. (In Chinese). Ye, Y.Q., Chen, G.J., Fan, H., 2003. Impacts of the “Grain for Green” project on rural communities in the Upper Min River Basin, Sichuan, China. Mountain Research and Development 23, 345–352. Yu, X., Zhang, X., Li, J., Zhang, M., Xie, Y., 2006. Effect of vegetation cover and precipitation on the process of sediment produced by erosion in a small watershed of loess region. Acta Ecologica Sinica 26, 1–8. Zhang, X., Zhang, Y., Wen, A., Feng, M., 2003. Assessment of soil losses on cultivated land by using the 137Cs technique in the Upper Yangtze River Basin of China. Soil and Tillage Research 69, 99–106. Zheng, F., 2006. Effect of vegetation changes on soil erosion on the Loess Plateau. Pedosphere 16, 420–427. Zhou, P., Nieminen, J., Tokola, T., Luukkanen, O., Oliver, T., 2006. Large scale soil erosion modeling for a mountainous watershed. In: Martin-Duque, J.F., Brebbia, C.A., Emmanouloudis, D.E., Mander, E. (Eds.), Geo-Environment & Landscape Evolution II. WIT press, UK, pp. 55–67. Zhou, P., Luukkanen, O., Tokola, T., Nieminen, J., 2007. Vegetation Dynamics and Forest Landscape Restoration in the Upper Min River Watershed, Sichuan, China. Restoration Ecology (OnlineEarly Articles). doi:10.1111/j.1526-100X.2007.00307.x. Zhou, P., 2008. Landscape-scale Soil Erosion Modelling and Ecological Restoration for a Mountainous Watershed in Sichuan, China. Viikki Tropical Resources Institute, Helsinki.