2.2.1. The InVEST Habitat Quality Model
An open-source Integrated Valuation of Ecosystem Services and Tradeoffs (InVEST) model of habitat quality version 3.6.0 (
https://www.naturalcapitalproject.org/invest/) has been developed recently at Stanford University by Sharp et al. [
32] to map and assess the habitat quality for individual LULC types [
32,
61]. The InVEST habitat quality model is a novel tool used for assessing habitat quality under anthropogenic threats [
37].
The key rationale for using this model include: (1) The model is relatively less data intensive and more flexible than other models; (2) it can easily be adapted to a specific context and readily available data either globally or locally [
14]; and (3) the model shows the hydrological and ecological connectivity developed by Vigiak et al. [
62]. Moreover, the unique characteristic of the InVEST habitat quality model is to analyze the spatial habitat quality trends and connectivity for each land-use type and quantify the sensitivity to threats.
For modeling the habitat quality, different geospatial data parameters were prepared using ArcGIS 10.4 (
Table 1). Based on Sharp et al. [
32] supported by other research such as Mashizi and Escobedo [
6] and Sallustio et al. [
63], used data on LULC, the relative weight of each threat (Wr), the habitat sensitivity of each threat (Sjr), the distance between habitats and sources of threats (Max. D), and habitat existence (presence/absence, Hj) to model habitat quality. These parameters were determined using expert knowledge supported by a field survey [
7].
Six woredas/districts and twelve villages/kebeles were selected for sampling using systematic random sampling (
Table 1) based on Tashakkori and Teddlie [
64]. To this end, intensive key informant interviews (KII) were undertaken in 12 villages in the districts with 72 experts (12 experts from each district, who had a good ecological understanding of the watershed) from the Guraghe zone, the environmental protection and natural resource office. The survey was designed according to the method used by Diehl et al. [
65] to better understand opinions regarding the importance of specific threats. The respondents were asked to identify the habitat existence/types, as well as the major threats found in the area: Water abstraction, sedimentation, population growth, pollution, invasive species, etc. The maximum distance of the threat affects the habitat quality and distance decay (the threat affects the habitat quickly or slowly) also responded by 72 experts. Next, opinions regarding the major threats’ impact on specific land-use types (forest, woodland, grazing land, shrubland, and cultivated land) were explored using the analytical hierarchy process (APH) model. The APH model was widely used to assign a weight to each threat by examining each threat’s potential impact on the habitat using a pairwise comparison matrix, with 1 to 5 preference scale [
66,
67] with the following Likert scale values: 5 (very high), 4 (poses a high threat), 3 (poses a medium threat), 2 (poses a low threat), 1 (poses no threat), and 0 (do not know). A matrix was built for each village by using expert-based judgment, containing threats that were used for the pairwise comparisons. Each element’s weight provided by an expert in the matrix was divided by its column total to generate a normalized pairwise matrix. Finally, the weighted matrix (averaging across the rows of the matrix) was generated using the eigen vector (priority vector), which shows each factor’s impact on the habitat [
68]. This is vital to assess pairs of threats as alternatives and compare each other in terms of their importance. The corresponding consistency ratio (CR) to verify the acceptable accuracy of pairwise comparisons in a judgment matrix is less than 10% [
66]. Finally, the average threat weight/sensitivity values for the 12 villages were used for each land-use type.
Focus group discussion (FGD) were also conducted for triangulation of the data and two clusters of FGD in each district/village were undertaken based on previous studies by Aneseyee et al. [
56]. The participants of the FGD, with range of 7 to 10 in one cluster (
Table 2) were elder, youth, and female representative, local administrator, and indigenous ecological knowledge person. Open and close ended structured questionnaires were prepared based on Arabatzis et al. [
69] to obtain data of habitat types, types of threats for the habitat, distance between threats, and the habitat and sensitivity of the impact to the habitat (see
Questionnaire S6 in Supplementary Materials). The fundamental points captured by KII with the experts and FGD include threats types in their surrounding habitat, causes for degradation of the habitat, estimating the maximum distance between the threats and habitat, the spreading rate of the threat (distance decay measurement), sensitivity measurement (which threat has more impact on the habitat), overall status of the habitat quality in their districts, and the types of measures used so far to control the threat for the ecosystem in the study area. Finally, the KII and the interviewees’ responses were collected and organized according to the InVEST habitat quality model requirement with respect to each LULC.
The model also assumes that the more sensitive a habitat type is to a threat (higher Sjr), the more degraded it becomes. In general, the impact of a threat on a habitat decreases as the distance from the degradation source increases and vice versa. Habitats closer to threats have suffered a higher impact, while threats far away from the habitat have a lower impact. Once the maximum effect distance of the threats and the threat impact (disturbance intensity) are determined by consultation, the spatial distribution of threats was determined using interpolation, Euclidian distance, and buffer in ArcGIS for the InVEST habitat quality model.
2.2.2. Data Preparation and Input for the InVEST Model
Cloud-free Landsat satellite images (
Table 3), with path = 169, row = 54, for the dry seasons (January to February) for 1988 (Landsat 5), 1998 (Landsat 5), 2008 (Landsat 7), and 2018 (Landsat 8) were obtained from the United States Geological Survey (USGS) data portal (
https://earthexplorer.usgs.gov). The spectral band spatial resolution used in this analysis was 30 m, and all the images of the investigated area were cloud free (0%). The images were already terrain-corrected and projected to the Universal Transverse Mercator (UTM) of WGS84 zone 37 N. The images were radiometrically calibrated to the reflectance value, and the Quick Atmospheric Correction algorithm was applied in the ENVI v5.2 software. The supervised image classification method was employed using the Mahalanobis distance classification. The reference data collected using GPS [
70] were first converted into a vector file, then into a region of interest (ROI) in the ENVI software. Using the ROI, the spectral signature of each LULC type has been extracted.
Due to the failure of the scan line corrector (SLC) encountered in 2003, the images acquired by the sensor had a data gap (striping) for 2008. Hence, we applied image gap-filling using a ‘gap-filling’ tool in ENVI v5.2 using two images, the acquisition dates of which were only two weeks apart.
A confusion matrix (error matrix) was used to measure the reliability of the LULC classification and validation [
71]. The matrix compared the ground true data obtained from reference sites with the classified image results for the sample areas. Thus, the error matrix yielded the overall accuracy, producer and user accuracies, and kappa statistic [
72].
Nine anthropogenic threats were identified in the watershed by following the approach of Terrado et al. [
7] and Wu et al. [
73]. They were agricultural expansion, invasive species, pollution, soil erosion, urbanization, population growth, water abstraction, and paved and unpaved roads (
Table 4). The watershed is one of the most densely populated areas of the country with a maximum estimated population density of 293 people per km
2 [
74]. What is more, the density follows an increasing trend (
Table 5). High population pressure has resulted in an increasing degradation of the natural environment, which disturbs the habitat quality [
75]. The population in the Gumer district of the watershed is higher and denser than in the five woredas (districts: Cheha, Ezha, Geta, Merab Azernet, Enemor) (
Table 5), which results in more significant habitat degradation. Moreover, the district cities affect biodiversity more than rural areas by generating waste.
The population data provided by the CSA [
59] and the 1987 forecast for other years were used with Equation (1) based on Haque et al. [
76]:
where
Nt is the forecast population size in a year;
P is the population size;
e is the natural logarithm to the base e (2.71828);
r is the rate of increase (natural increase divided by 100);
t is the period involved.
Urbanization increased by 444.73% from 891.23 ha in 1988 to 3963.6 ha in 2018 (
Figure 3). Urbanization includes dense settlements, commercial areas, multiple governmental and private institutions, schools, religious institutions, health centers, educational institutions, and other social service areas. Hence, these institutions could generate high waste that is considered a threat to biodiversity and have irreversible environmental effects that affect the quality of the surrounding habitat. However, distance to urbanized areas was considered an important determinant factor for habitat quality rating. Accordingly, the experts recognized that urbanization has a maximum impact of 7 km and it has less impact on the habitat quality at a large distance than at a proximate distance.
According to a Guraghe municipal annual report on waste (m3/month) for Ezha (1624 m3), Cheha (1682 m3), Gumer (1598 m3), Merab Azernet (1534 m3), Gunchre (1512 m3), and Geta (1489 m3), waste generation affects habitat quality up to 20 km causing serious pollution of water bodies, soil, dumping of nonbiodegradable plastic to pristine environment and unpleasant odor. The solid waste is also promoting wildlife and human disease transmission by creating a favorable condition for reproduction of harmful insects and reptiles, bacteria and virus, which might be damaging the grazing land and the vegetation. Dumped solid waste also causes car accidents on the road, illegal open dumps often block the drains and sewers. Strong wind and storms are spreading dust and filth from the open dumps of solid waste to adjacent areas.
Agricultural expansion (
Figure 4) is also a threat to biodiversity due to the destruction of the habitat. Vegetation clearing is also likely, hindering the regulating capacity of the land, which results in hampered biodiversity and affects habitat occurrence up to 1 km.
The InVEST sediment delivery model was used to simulate the potential soil erosion in the watershed based on the RUSLE equation of Wischmeier and Smith [
77], using six factors (Equation (2)):
where
is the mean annual soil loss (t/ha/year),
R is the rainfall erosivity (MJ Mm/ha/year),
K is the soil erodibility (MJ mm/t/ha),
LS is the topographic factor (dimensionless) (length and gradient), which was generated from DEM,
C is cover (dimensionless) simulated by means of normalized difference vegetation index (NDVI), and
P is management (dimensionless).
The erosivity factor was calculated using Equation (3), as provided in Hurni [
78] and derived from spatial regression analysis for Ethiopian conditions, based on the available mean annual rainfall (P) for thirty-year data of daily and monthly rainfall from meteorological stations in the watershed (Emdiber, Gubre, Gunchre, Agena, and Gumer). Inverse distance weighted interpolation (IDW) was used to create the spatial raster map. The interpolated R factor was used as input for the InVEST model for the year 1988, 1998, 2008, and 2018:
where R is the erosivity factor and P is the mean annual rainfall in mm/year.
Based on the Food and Agriculture Organization (FAO) (1984), a soil map of Ethiopia (1:2,000,000), was generated and the shapefile of the legacy soil map was obtained from the FAO and complemented by the Omo-Gibe River Basin master plan soil database, made available by the Ethiopia Water, Irrigation and Electricity Ministry (MOWIE). Five soil types were identified in the study area; Chromic Luvisols, Pellic Vertisols, Eutric Nitosols, Leptosols, and Orthic Acrisols. In each soil type, plot soil samples were taken to determine the K-factor using Equations (4) and (5), based on the four soil characteristics that determine erodibility (sand, clay, silt, and soil organic matter content). The organic matter and soil texture were analyzed at Wolkite soil laboratory by taking 20 soil samples, considering each LULC, from the five soil types (total of 100 samples) of the watershed.
where K is soil erodibility, OM is soil organic content (%), and
The corresponding average C-value was determined for each LULC class for the years (1988, 1998, 2008, and 2018) based on NDVI value using Equations (5) and (6), following Durigon et al. [
79]:
where NIR is the surface spectral reflectance in the near-infrared band and RED is surface spectral reflectance in the red band.
The NDVI was computed by processing Landsat images for 1988, 1998, 2008, and 2018 in the ENVI v5.2 software and ArcGIS spatial analysis was used to extract NDVI for each LULC in the reference years. A higher NDVI value indicates higher vegetation cover and vice versa. Moreover, NDVI is determined by a combination of bands based on Landsat images but the formula for using bands over a period of time varies. The 2018 NDVI formula is different from the rest of the years. LANDSAT 5–7 used NDVI = (B4 − B3)/(B3 + B4) for 1988, 1998, and 2008. For 2018, NDVI = (B5 − B4)/(B5 + B4), was used by Landsat 8. Landsat 5 is used for 1988 and 1998, Landsat 7 is used for 2008, and Landsat 8 is used for 2018.
Management practices (P) were determined based on field observation in the upper, middle, and lower watershed, depending on the different conservation practices, and the value of the P-factor ranges from 0 to 1, as suggested by Ganasri and Ramesh [
80]. For agricultural lands, the p-value was assigned based on slope classification, using a reclass tool in the ArcGIS software. The mean annual soil loss in the watershed increased from 6.03 t/ha/year in 1988 to 17.91 t/ha/year in 2018 (
Figure 5). Soil erosion results in decline in soil fertility in upstream parts of the land affecting farmland at a maximum distance of 2 km. The erosion-based constraint significantly affects the downstream ecosystem degradation such as sedimentation of reservoirs created with dams, the burial of fertile agricultural fields with boulders and stone, etc.
The threat data on invasive species, pollution, water abstract, and road system were not found for 1988, 1998, and 2008. Therefore, the same data were used for 1988, 1998, 2008, and 2018. The spatial distribution threats effect of invasive species and pollution were mapped using IDW interpolation techniques of the ArcGIS whereas road was spatially mapped using buffer and water abstraction delineated using Euclidean distance.
Invasive species are considered a threat to biodiversity, as well due to the possibility of a quick invasion on productive land and their allelopathic effect on the surrounding biodiversity. The authors investigated the density of invasive species by systematic field sampling and count on 20 by 20-m patches on 60 plots. The invasive species were identified by a botanist and using a field manual. They were
Senna didemobotrya (50 stems/ha),
Parthenium hysterophorus L. (100 stems/ha),
Argemone ochroleuca (70 stems/ha),
Calitropis procera (80 stems/ha), and
Nicotiana glauca (wild tobacco) (48 stems/ha). They were more prevalent on the roadside and near urbanized areas (
Figure 6E).
There are three private water abstraction facilities producing drinking water in the watershed, Aden, Fiker, and Waw. They were established after 2012. The amount of water abstracted per year in Aden, Fiker, and Waw was 100, 67, 50 million m3, respectively. This water abstraction affects the ecosystem of the watershed by reducing available water. It further has a significant impact on the downstream Omo-Gibe dams. Therefore, the greater the distance from the water production industry, the greater the persistence of biodiversity in the area and the lesser the sensitivity to the threat.
The primary road system has high traffic density (
Figure 6A). It becomes a threat to biodiversity because of exhaust gas pollution, soil pollution, noise pollution, and the danger to animals that cross the street. The secondary road system (
Figure 6B) also influences the habitat quality by providing transport opportunities for people and animals even though its impact is lower than that of the primary road system. The road system also fosters the timber and nontimber forest product harvesting, which leads to deforestation and disturbance of the habitat quality. Therefore, a primary road is considered a threat to the maximum distance of 1.5 km and a secondary road is assumed to be a threat at a distance up to 0.5 km. The absence of any roads prevents degradation of biodiversity.
The relative habitat suitability score (intensity of the threat) can be assigned to each LULC type, ranging from 0 to 1, where 1 indicates the highest habitat suitability and 0 indicates the lowest suitability [
36,
48]. Threats and sensitivity scores were determined based on literature data and expert knowledge (
Table 6).
The InVEST model is used as a half-saturation curve to show the relationship between habitat degradation scores and habitat quality scores [
81]. Threat decay in space is described as either a linear (Equation (8)) or exponential (Equation (9)) distance-decay function. The InVEST model determines the impact of threat r that originates in grid cell y,
ry, on habitat in grid cell
x as
irxy in Equation (8):
where
dxy is the linear distance between grid cells x and y and
dr max is the maximum effective distance of threat r’s reach in space.
The total threat level in grid cell
x with LULC or habitat type
j is determined with Equation (10):
where
r is the LULC threat, with
r = 1, 2,…n;
R is the index of all the modeled degradation sources;
y is the index of all grid cells on r’s raster map;
Yr indicates the set of grid cells on r’s raster map;
wr are the weight parameters;
βx is the accessibility factor.
The quality of habitat (Q
xj) in parcel
x or LULC j is calculated with Equation (11):
where
Hj is each LULC assigned a habitat score from 0 to 1 (1 = highest habitat suitability, 0 = no habitat),
z = 2.5 and
k are scaling parameters (or constants). The
k constant is 0.5.
The LULC area for 1988, 1998, 2008, and 2018 has been drafted (
Table 7). The land has been classified into eight classes with a numeric LULC code for each class: Bare land (1), built-up land (2), shrubland (3), cultivated land (4), forests (5), grazing land (6), water bodies (7), and woodland (8). The LULC raster encompasses the area of interest and a buffer the width of which is the greatest maximum threat distance.