Study area
This study was conducted at Gojeb sub-basin of Omo-Gibe Basin in the southwest Ethiopia (Fig. 1). The Gojeb River catchment (a tributary of the Omo River) is covering an area of about 700000 ha. Geographically, the catchment is located between 36.16 & 37.492 East and 07.12 & 08.13 North with altitudinal range of 806 to 3348 m.a.s.l. The catchment lies in two regional states, i.e., the Southern Nations and Nationality Peoples Region (SNNPR) and Oromia Region. Climate of the study area is generally classified as tropical cool humid. The agroecology of the catchment consists of cold moist-Dega (around the upper catchment of Gojeb River), hot moist - Kola (the middle portion of the southern part of the catchment) and wet moist -Woina Dega (the remaining substantial part of the catchment). Annual rainfall varies from about 1000 mm in the extreme south to over 1850 mm in the highland northern parts of the catchment with the average being over 1450 mm (Yilkal, 2019).
The study area covers about 700000 ha in the Gojeb sub-basin in the Omo-Gibe-Basin (Fig. 1). Gojeb River is partly bordering Oromia and SNNP regions. The detailed analysis of land transformation of the watershed was identified more than 25 land use/land cover types (Table 1) as small as 0.03 ha and as large as 269499 ha of land with 30 m spatial resolution and within 30 years’ time span. These detailed land use land cover classes were also summarized into commonly used corresponding land use land cover types.
Land use datasets and approaches
Landsat 30 m satellite images analyses were conducted between 1986 and 2016 to monitor 30 years land transformation using the advantages of remote sensing and geographic information system. They provide wide ranges of opportunity to produce LULC data at various scales. However, generating complete and comprehensive LULC information via remotely sensed data that fill the wide range of the need still faces difficulties (Jensen, 1996; Renison et al., 2004). In Ethiopia, two major categories of challenges limit the potential of remote sensing techniques to produce the required scale and accuracy of LULC information: (1) landscape complexity (topographic and farming system) and (2) accessibility of better resolution remotely sensed data and suitable classification approach (Kassawmar et al., 2016a). Experiences shows that a stratified mapping approach can potentially address the challenges encounter when mapping heterogeneous and large areas (Homer et al. 2000; Lu et al., 2015).
Deriving Homogenous Image Classification Units (HICUs) was used that subdivided each Landsat image into smaller units where similar land cover mosaics occur was found to be a suitable classification approach. HICU development was done using multiple information sources such as altitude, terrain, farming system, rainfall pattern and soil. This approach resulted in a varying amount of HICUs within a Landsat image, depending on the location and landscapes types, followed by identifying the dominant and subordinate land cover features for each HICU, or majority and detailed classes. These land features into were grouped detailed and majority classes based on the occurrence, dominance and distribution of the land features. This requires disaggregation of classes at multiple steps of classification, which could allow capturing smaller classes that are commonly ignored in large and complex landscape mapping. For example, the class forest generated from church forests and protected high forests, which are found either in high moist area or dryland forest. Finally, manual digitizing, edge enhancement and NDVI thresholding were used to extract the majority and/or detailed classes from the imagery. The extracted classes were combined and areas of their occurrence masked within each Landsat image so that they would not distort the further classification process (Table 1).
Different approach has shown a considerable improvement on the accuracy of classification in complex landscapes. However, accuracy of classification could be more improved based on the number and precision of segments developed for every scene and local knowledge of the area (Kassawmar et al., 2016). In complex landscapes traditional classification techniques that apply full scene as unit of analysis have shown limitations due to spectral variability of features related to bio-physical complexity observed in a scene (Kassawmar et al., 2016b). Then major classes were generated from the detailed (Level II) classification result carried out from Landsat image with corresponding ground truth and ancillary data with an accuracy assessment obtained > 85% (Kassawmar et al., 2016).
Table 1
Descriptions of detailed LULC class names (Level II) and corresponding LULC names (Level I)
Class name
|
Description
|
LULC name
|
Class name
|
Description
|
LULC name
|
Afcf
|
Agroforestry dominated by coffee
|
Cropland
|
Hghf
|
High forest
|
Forest
|
Afec
|
Agroforestry enset & coffee
|
Cropland
|
Homp
|
Home garden planation
|
Plantation
|
Afen
|
Agroforestry dominated by enset
|
Cropland
|
Mixf
|
Mixed forest
|
Forest
|
Agrf
|
Agroforestry
|
Cropland
|
Pfor
|
Plantation forest
|
Plantation
|
Bare
|
Bareland
|
Bareland
|
Rivc
|
River courses
|
Forest
|
Crhil
|
Crop in the hillside
|
Cropland
|
Rivf
|
Riverine forest
|
Forest
|
Crwt
|
Crop with trees
|
Cropland
|
SBdn
|
shrub-bush dense
|
shrubland
|
Csht
|
Cultivated with shifting
|
Cropland
|
SBop
|
Shrub-bush open
|
Shrubland
|
Cwot
|
Crop without trees
|
Cropland
|
SeTT
|
Settlement
|
Settlement
|
Dghi
|
Degraded hills
|
Shrub
|
Swmp
|
Swamp Waterbody
|
Dryf
|
Dry forest
|
Forest
|
Wate
|
Waterbodies
|
Waterbody
|
Gdry
|
Grassland drained
|
Grassland
|
Wldn
|
Woodland dense
|
Woodland
|
Gsvn
|
Grassland savanna
|
Grassland
|
Wlop
|
Woodland open
|
Woodland
|
Gwet
|
Grassland wet
|
Grassland
|
|
|
LULC changes were calculated for two different time periods: for 1986 and for 2016, and the change between 1986 and 2016 using cross-tabulation (Kindu et al., 2016; Shiferaw et al., 2019) and calculating percent changes for each LULC type over time (Gashaw et al., 2018; Temesgen et al., 2018). Furthermore, class-specific gains, losses and stable areas, as well as total change area and net changes of the total area analyzed were calculated (Alo and Pontius, 2008; Zewdie and Csaplovics, 2015). Finally, annual change rates were calculated for each LULC type following Puyravaud (2003) and Tilahun et al. (2014), i.e. the rate of change for a specific class was calculated by dividing the class-specific changes between two time intervals by the number of years between these two observed points in time. LULC changes were calculated for two different time periods between 1986–2017 as methods applied by different studies (Eckert et al., 2017; Kindu et al., 2016; Shiferaw et al., 2019) and calculating percent changes for each LULC type over time (Gashaw et al., 2018; Kindu et al., 2016; Temesgen et al., 2018) (Eq. 1).
Where, A1 is area of land use and land cover type (ha) in year 1, A2 is area of land use and land cover type (ha) in year 2.
Furthermore, class-specific gains, losses, and stable areas, as well as total change area and net changes of the total area analyzed were calculated (Alo and Pontius, 2008; Zewdie and Csaplovics, 2015). Finally, annual change rates were calculated for each LULC type following Puyravaud (2003) and Tilahun et al. (2014), i.e. the rate of change for a specific class was calculated by dividing the class-specific changes between two time intervals by the number of years between these two observed points in time (Eq. 2).
where, A1 is area of land use and land cover type (ha) in year 1, A2 is area of land use and land cover type (ha)in year 2, Z is the time interval between A1 and A2 in years.
Ecosystem serving values
In this study, the benefit transfer approach was used to estimate ecosystems service values (ESVs) of different LULC types and their changes (Costanza et al., 1997, 2014; Niquisse et al., 2017). The benefit transfer approach refers to the process of using existing values and other information from the original study site to estimate ESVs of other similar locations in the absence of site specific valuation information (Bagstad et al., 2013; Niquisse et al., 2017). We calculated the ESVs of the LULC types in Gojeb catchment by adapting the coefficients of tropical areas on regional estimates of ESVs using data provided by Kindu et al. (2016), who conducted a study on LULC and ESVs in Ethiopia using conservative estimates of ESV coefficients, which were based on values from studies conducted in areas similar to the geographical setting of our study area. These ESVs include the main three ES: supply, regulation/monitoring and provision (Kindu et al., 2016) and also using the updated global coefficients provided by Costanza et al. (2014). Land use types such as bare land and settlement did not have a coefficient in some studies (Costanza et al., 1997; Kindu et al., 2016; Tolessa et al., 2017). Hence, the ESVs for all LULC types were calculated for each period using the following Equation (Costanza et al., 1997, 2014) and similar to studies conducted for tropical forests provided by Costanza et al. (2014) and that for woodland/shrub land provided by Temesgen et al. (2018) (Eq. 3):
Where, ESV is estimated ecosystem service value, Ak is the area (ha) of LULC type k, and VCk is the value coefficient (US$ ha − 1 yr − 1) for LULC type k.