Study area
The study was conducted in Bursa and Wonsho districts of Sidama Zone, South Nations Nationalities and Peoples Regional State (SNNPRS) of Ethiopia. The study area covers about 2800 ha comprising three Kebeles (i.e., the lowest administrative unit in Ethiopia) in two adjacent watershed areas, namely Lalta and Kolla. Two of the three Kebeles, Mollona-Meno and Chulule, are located in Bursa District, whereas the third kebele known as Orowo is found in Wonsho District. All the three Kebeles are assumed to represent the Ethiopian highlands in sub-afro-alpine vegetation Zone in general and the Sidama highlands in particular (Fig. 1). The study area is located between latitudes of 6° 41′ 19″ and 6° 44′ 52 ′′ N, and longitudes of 38° 33′ 52″ to 38° 35′ 59″ E at an altitude range between 2749 and 3172 m above sea level.
The topographies of the study areas are mountainous and hilly, with slope percentages ranging from 10 to 45%. The mean annual rainfall is about 2000 mm, whereas the mean annual temperature varies between 10 and 15 °C. The area receives bi-modal rainfall with the short rainy season or the Belg rain occurring from March to May and the main rainy season or the Kremt rain occurring from June to October (Bursa District Ministry of Agriculture office (MoA) 2014).
The study areas are occupied by low forests and grasslands. The vegetation is characterized by trees, bushes, and shrubs. Some of these included Erica arborea, Hagenia abyssinica, Ilex mitis, Hypericum revolutum, Juniperus procera, Yushania alpina (the giant grass/bamboo), Rubus steudneri, Discopodium penninervium, Dombeya schimperiana, Vernonia auriculifera, and Phytolacca dodecandra.
Sampling and survey approach
The survey was carried out in February 2016. A total of 240 household (HH) heads, i.e., 80 HH heads per Kebele, were selected purposively for individual interview. Only seven female respondents were represented in the present survey due to few female-headed HHs in the study area. Age was one of the criteria in the selection of HH heads for the interview. We selected HH heads aged 55 years old and above so that they can easily recall and tell trends of environmental changes over the last 40–50 years. The survey was conducted in two phases. Firstly, field observations were undertaken to understand the different classes of land use and gather information on the background of the study areas for instance farming systems such as crop farming and livestock production. This was followed by preparation and pre-testing of a semi-structured questionnaire, which consists of a mix of unstructured and structured questionnaires (i.e., questionnaire that contains both open-ended and closed-ended questions). Secondly, the information gathered during the first phase of the field visit was used to redesign the semi-structured questionnaire to collect qualitative and quantitative data. Semi-structured face-to-face HH heads interviews were conducted to gather information on trends, factors facilitating land use/land cover changes, and indicators used by the local farmers to tell information on climate change and its impacts, as well as current land management practices and socio-economic conditions of the local communities in the mountain grasslands of Sidama Zone. In this paper, drivers of land use change refers to climate- and human-related factors that directly/indirectly impact the processes of land use change and associated resources, such as soil, water, and biodiversity, while local indicators refers to farmers way of understanding changes in climate and land use changes. The questionnaire covered a wide range of socio-economic conditions that included (1) land size, (2) land-use types and trends of environmental changes over the last 40–50 years, and (3) communities’ perceptions on drivers of land use/cover changes. Generally, data were kept safe and secure at all stages of research process and entered into a computer for data analysis.
Data analysis on households interview
We used Statistical Package for Social Sciences (SPSS) version 20 (SPSS Inc. 2008) software to analyze the household survey data. Descriptive statistics such as cross-tabulation, frequencies, and percentage values were employed to summarize the qualitative data set.
Specifications of satellite data and preparation
The data used in the current study were classified into satellite data and data based on ground truth in terms of land use classes. The ground truth data were in the form of reference data points collected using global navigation satellite system (GNSS) in February 2016 for the 2015 map of land use/cover changes. Satellite image data for a period of 4 years that consisted of multi-spectral scanner (MSS) data acquired by Landsat satellite 1 on 31/01/1973, Thematic Mapper (TM) data acquired by Landsat 5 on 21/01/1986, Enhanced Thematic Mapper plus (ETM+) data acquired by Landsat 5 on 05/02/2000, and operational land imager (OLI) and thermal infrared sensor (TIRS) data acquired by Landsat 8 on 23/12/2015 (Fig. 2). All data were downloaded from “Global Land Cover Facility” (http://glcfapp.glcf.umd.edu:8080/esdi/) and “Earth Explorer” (http://earthexplorer.usgs.gov/) websites. Specifications of the satellite data acquired for change analysis are given in Appendix 1.
Image pre-processing and classification
Collection of different episodic years of satellite image data was the initial stage for image processing and analysis. Multi-temporal rectified Landsat images were downloaded from “Global Land Cover Facility” and “Earth Explorer” websites. Landsat satellite images of the year 1973, 1986, 2000, and 2015 having a map projection of Universal Transverse Mercator (UTM) zone 37 and datum WGS84 were pre-processed using Earth Resources Data Analysis System (ERDAS) Imagine 2014 software in order to well-match the image with other ancillary data (Fig. 2). Satellite image register an exhaustive data of features on the ground at the time of data acquisition. As a pre-processing stage, layer stacking (compiling of different bands) and sub-setting were done on the different images. Then, digital image enhancement and interpretation techniques were used to make the interpretation of the data easier. The main image processing started with selecting sample sites using the field GNSS data and all the available images were classified into five classes (i.e., forestland, grassland, shrubland, bushland, and agricultural land) by applying supervised classification method and maximum likelihood algorithm. After the land use/cover changes classification of the years 1973, 1986, 2000, and 2015, the classification results were evaluated by employing accuracy assessment technique to examine how the results reflect the reality on the ground. Besides, the conversion matrix analysis was conducted to investigate the source and destination of land use/cover changes. Descriptions of land use/land cover changes are given in Appendix 2.
Land change detection
A post-classification detection method was employed to perform land use/land cover change detection. According to Singh (1989), change detection is the process of identifying differences in the state of an object or phenomenon by observing it at different times. Lu et al. (2004) indicate that detection of changes in land-cover patterns is extremely important for understanding relationships and interactions between human and natural phenomena in order to make decision concerning the management and conservation of the natural environment. A pixel-based comparison was used to produce change information on pixel basis and thus, interpret the changes more efficiently taking into account the advantage of “-from, -to” information. Classified image pairs of two different decades’ data were compared using cross-tabulation in order to determine the qualitative and quantitative aspects of land use change between 1973 and 2015. A change matrix (Weng 2001) was produced with the help of ERDAS Imagine software. Quantitative areal data of the overall land use/cover changes as well as gains and losses between 1973 and 2015 in each category were compiled. Maps were paired (1972–1986, 1986–2000, and 2000–2015) and overlaid in a geographic information system (GIS) in order to produce three cross-tabulation matrices that combined classes among the three pairs of periods to produce areas.
The diagonal of each matrix showed the area of land that remained unchanged during each period: this is the so-called persistence in the landscape. The off-diagonal combinations account for the patches that were transformed from one category into another. Gain, loss, absolute value of net change, swap, and total change were calculated for each class according to the method of Pontius et al. (2004). The gains are the differences between the column totals and persistence while the losses are between row totals and persistence. With respect to loss-to-persistence ratios (lp), values higher than 1 indicate a greater tendency of transition to other land use land cover (LULC) class rather than stability, which is a measure of the vulnerability to transition (Braimoh 2006). Gain-to-persistence ratio (gp), if a value is higher than 1, it indicates a greater tendency of LULC types to gain rather than to persist (Braimoh 2006). The absolute value of net change was, therefore, the modulus of the difference between time 1 and time 2. Swap was a component of changes that imply that a given area of a category is lost at one location, while the same area is gained at a different location. Swap was calculated as twice the minimum of the gain and loss, since each gained pixel is paired with a lost pixel to create a pair of pixels that swapped. Finally, the total change was either the sum of the net change and the swap or the sum of the gains and losses. Gain-to-persistence ratio (gp = gain/persistence), loss-to-persistence ratio (lp = loss/persistence), and net change-to persistence ratio (np = gp − lp) were also calculated in order to evaluate the persistence characteristics of the different LULC classes (Braimoh 2006). Systematic transitions were identified by comparing off-diagonal changes to the expected values of change, considering random processes of gain or loss. A process of gain was considered random if the gaining category replaced others proportionally to the area of those other categories at time 1. Therefore, a category that gained would randomly replace a larger percentage of a large category and a smaller percentage of a small category. Equation (1) gives the expected transition from class i to class j (Gij), where P •j is the column total of category j in the transition matrix and denotes the percentage of the landscape in category j at time 2, Pjj is the persistence of category j, Pi• is the row total of category i and represents the percentage of the landscape found in category i at time 1, and, finally, Pj • is the row total for category j, which denotes the percentage of the landscape in category j at time 1.
$$ \mathrm{Gij}=\left(\mathrm{P}\bullet \mathrm{j}-\mathrm{Pj}\mathrm{j}\right)\times \mathrm{Pi}\bullet /100-\mathrm{Pj}\bullet $$
(1)
Conversely, a process of loss is random when the losing category is replaced by others proportionally to the area of those other categories at time 2. In this case, a category that loses would be replaced randomly to a greater extent by a large category and to a lesser extent by a small category. Equation (2) gives the expected transition from class i to j (Lij), where Pi• is the row total of category i, Pii is the persistence of category i, P •j is the column total of category j, and P•i is the column total for category i.
$$ \mathrm{Lij}=\left(\mathrm{Pi}\bullet -\mathrm{P}\mathrm{ii}\right)\times \mathrm{P}\bullet \mathrm{j}/100-\mathrm{P}\bullet \mathrm{i} $$
(2)
Any large deviation from the expected values was considered to represent a systematic landscape change. Finally, the fact that class X systematically gained from class Y, while class Y systematically lost to class X, gives conclusive evidence of a dominant signal of landscape transformation (Braimoh 2006). The classified images were compared in three periods, i.e., 1973–1986, 1986–2000, and 2000–2015. Change statistics were computed by comparing image values of one data set with the corresponding value of the second data set in each period. LULC conversion matrix between 1973 and 2015 was generated using ArcGIS 9.1 software and compiled in a matrix table, and the values were presented in terms of hectares.
Accuracy assessment
Accuracy assessments determine the quality of the information derived from remotely sensed data. The product of image classification is land cover maps. Their accuracy needs to be assessed so that the ultimate user is made aware of the potential problems associated with their use. Evaluation of the accuracy of a classification may be undertaken for each of the categories identified and its confusion with other covers, as well as for all the categories. The outcome of accuracy assessment is usually presented in a table that reveals accuracy for each cover category and for all categories as a whole.
Assessment of classification accuracy of the different images between 1973 and 2015 was carried out to confirm the quality of information derived from the data. According to Owojori and Xie (2005), if the classification data are to be useful in detecting change analysis, it is essential to perform accuracy assessment for individual classification. For the accuracy assessment of land cover maps extracted from satellite images, stratified random sampling method was used to represent the different land cover classes in the study areas. The accuracy assessment was carried out using 70 random points for each class, which was based on ground truth data and visual interpretation. Comparison of reference data and classification results were carried out statistically using error matrices. In addition, a non-parametric Kappa test was also performed to measure the extent of classification accuracy as it not only accounts for diagonal elements but also for all the elements in the confusion matrix (Rosenfield and Fitzpatirck-Lins 1986). Kappa is a measure of agreement between predefined producer ratings and user assigned ratings. It is calculated by the formula: K = P(A)P(E)/1P(E), where P(A) is the number of times the K raters agree, and P(E) is the number of times the K raters are expected to agree only by chance (Gwet 2002; Viera and Garrett 2005).