Historical and future land use and land cover changes in the Fincha watershed, Ethiopia


 The increasing human pressure on African regions is recognizable from land use land cover (LULC) changes maps, as derived from satellite imagery. Using the Ethiopian Fincha watershed as a case study, the present work focuses on i) identifying historical LULC change in the period 1989-2019; ii) estimating LULC in the next thirty years, combining Geographical Information Systems (GIS) with Land Change Modelling (LCM). Landsat5/8 images were combined with field evidence to map LULC in three reference years (1989, 2004, 2019), while the Multi-Layer Markov Chain (MPL-MC) model of LCM was applied to forecast LULC in 2030, 2040 and 2050. The watershed was classified into six classes: waterbody, grass/swamp, built-up, agriculture; forest and shrub. The results have shown that, in the past 30 years, the Fincha watershed experienced a reduction of forest and shrubs due to ever-increasing agricultural activities, and such a trend is also expected in the future. In addition, the decrease in areas covered by natural forests can drive to an increase in soil erosion, fostering the siltation in the water reservoirs located in the basin. The study pointed out the urgency of taking actions in the basin to counteract such changes, which can eventually drive to a less sustainable environment.


Introduction
Land use is de ned as how the land is utilized by human beings and their habitats, usually with an accent on a practical role of land for economic activities, whereas land cover is a physical characteristic of the Earth's surface or attributes of a part of the Earth's land surface and immediate subsurface, including biota, soil, topography, surface and groundwater, and human structures (Alan et al. 2020;Leta et al. 2021a;Regasa et al. 2021;Tadese et al. 2021). As it strictly connected with representing the hydrological cycle (Dwivedi et al. 2005), land use and land cover (LULC) change has been one of the most widely used methods to comprehend past land uses, types of changes estimated, the forces behind such changes and the perceptible transformations of the Earth's surface (Alam et al. 2020). LULC changes could involve critical issues such as biodiversity degradation and negative impact on human life (Kenea et al. 2021;Khan et al. 2021).
The study of LULC change has attracted growing interest in recent years, as it is a complex issue that involves physical, environmental, and socioeconomic facts. According to Lambin et al. (2000), the modelling of land cover processes can answer questions like i) which are the main environmental and cultural variables that contribute most to the observed changes, and why? ii) within a geographical region, which location can be affected by land cover changes, where and at what rate do land cover change and when? Prediction using time serious data is important for the future management plan of LULC, and it is frequently employed for a diverse appropriateness measure as a proxy of human in uence on land change processes (Han et al., 2015;Tadese et al. 2021). Analysis of the historical trends of LULC is paramount in modelling future LULC, as past information are generally representing a good proxy of human in uence on land processes (Han et al., 2015;Rahman et al., 2017). To adequately predict future scenarios, models should be calibrated and validated, and various techniques are available in the literature. Geometric rectifying, supervised and unsupervised classi cation methods, post-classi cation method, GIS spatial analysis, Markov chain analysis, ERDAS imagine model and Land Change Modeler (LCM) were used to analyze historical LULC change and predict future changes (Fan et al. 2008).
The Land Change Modeler is a model used to evaluate the changing trend from one land use category to another and has been found to provide high spatial and temporal resolution with a reduced computational effort (Etemadi et al. 2018;Leta et al. 2021b). The integrated CA-Markov model is a robust technique in terms of quantity estimation as well as spatial and temporal dynamic modelling of LULC, because remote sensing data and GIS can be pro ciently incorporated. Integrated CA-Markov model can translate the results of the Markov chain model to spatially explicit results (Al-sharif et al. 2014). The Markov model has been widely adopted in ecological studies, and its applicability in LULC change modelling is promising because of its ability to quantify not only the states of conversion between land-use types, but also the relative rate of conversion (Subedi et al. 2013).
Few studies of LULC change provided a combined assessment of the driving forces and consequences of such variations, particularly in Africa (Reid et al. 2000). Recent research focused on sub-Saharan Africa, showing a decrease of land covered by the natural environment, mostly due to human activities such as population growth, economic development, and globalization (Mussa et a. 2017;Näschen et al. 2019). However, only a few studies tried to explicitly connect such changes with the loss of ecosystem services (Tolessa et al. 2017).
In this region, the dynamics of LULC intensities and rates are changing and highly associated with overexploitation of natural resources, while the process is governed by climate (long dry periods followed by heavy precipitation), soil characteristics (thin layer of topsoil, silty texture, or low organic matter content), vegetation (barren land), topography (steep-slope), and natural hazards (forest re, landslides) (Nasir and Selvakumar, 2018;Leta et al. 2021b). During the last decades, the human pressure increased signi cantly, therefore the process has been accelerated considerably. In Ethiopia, LULC changes are persistent events where agricultural activities and settlements are dominant in the rural landscape (e.g., Alemu et al. 2015;Dibaba et al. 2020;Hailu et al. 2020). For example, Tadese et al. (2021) reported that the agricultural and settlement increased by around 17% and 3%, respectively, from 1987 to 2017 while agricultural land decreased by 78%.
Focusing on the Fincha River basin in Ethiopia, this research aims to: i) understand the historical LULC changes in the past three decades ; ii) predict possible LULC patterns in the future three decades . The objectives of the paper will be tackled by combining satellite imagery (Landsat dataset) and modelling (LCM), using eld evidence to support the study. The study area and the used methods will be described in Section 2, while the results are presented and discussed in Section 3.
Four main seasons are characterizing the region: Summer, from June to August, with heavy rainfalls; Autumn from September to November, called harvest season; Winter, from December to February, is the dry season with frost in the morning especially in January; Spring, from March to May, with occasional showers, is the hottest season. The annual rainfall of the study area ranges between 1367 and 1842 mm with the minimum rainfall occurring in the Northern lowlands and maximum rainfall greater than 1500mm rain in the Southern and Western highlands. From June to September is the main rainy season of the catchment, with an average of 1604 mm and the maximum rainfall between July to August. Natural resources such as the Fincha, Amarti and Nashee lakes contribute to the national economy by generating hydroelectric power, but are also used for irrigating large elds devoted to sugar cane. The area is of interest for national and international hydro-politics due to its downstream connection to the Nile basin and the intense agriculture.

Dataset
The study was performed using freely available satellite imagery and a Digital Elevation Model (DEM). The latter, having a resolution of 30m and referring to 2019, was acquired from the GIS and Remote Sensing Department, Ministry of Water, Irrigation and Energy of Ethiopia.
Landsat-5 TM (L5, for the years 1989 and 2004) and Landsat-8 OLI-TIRS (L8, for the year 2019) data were downloaded from the United State Geological Survey (USGS) website (earthexplorer.usgs.gov). The images referred to January, when there is a clear sky corresponding to the dry season, and were atmospherically corrected via QGIS (qgis.org). To cover the whole watershed area, a composite of Landsat images from different paths/rows was created, assuring that the images refer to the same season (Table 1). Field surveys have been conducted to assist the LULC classi cation of the satellite images. In addition, key informant interviews (KII) and focal group discussions were performed to obtain socio-economic support data, as this is paramount to understand how locals interact with the environment (Betru et al. 2019;Dibaba et al. 2020). KII were conducted with elders, as they have known the area for at least 30 years and had good knowledge on past LULC changes Focal group discussion was conducted with experts from zonal and district o ces of Agriculture, Natural resources management, Environment and climatic change, Land use administration and local people's representatives. Ground truth data were collected using GPS and digital cameras to evaluate the current LULC.

Land use land cover classi cation of historical data
To map LULC, satellite images should be classi ed, assigning prede ned LULC classes to some pixels. As pointed out by Jemberie et al. (2016), this phase could be affected by various factors such as classi cation methods, algorithms, collecting of training sites, and the quality (correctness) of the classi cation should be assessed via eld evidence (Kindu et al. 2013;Abijith and Saravanan, 2021).
The study was performed classifying three reference years (1989,2004,2019), and accounting for six classes, as summarized in Table 2. The maximum likelihood supervised classi cation method was applied via ArcGIS by creating training sites signature.
Training sites for the L8 image of 2019 were de ned using 100 ground truth points, while, for the two older L5 images, training signature sites were de ned via unsupervised classi cation, ancillary data (Google Earth), KII information and literature data (Kenea et al. 2021). To improve image quality, quality assessments were used by taking a total of 50 ground truth points (20 agriculture, 5 waterbody, 5 built-up, 10 forest, 5 shrub, 5 grass/swamps).
To quantitatively assess the accuracy, statistical methods like overall accuracy and kappa value were applied. Based on this, random sampling data's were prepared to check the overall accuracy OA and to determine the Kappa coe cient K. Comparing the total corrected samples TCS and the total samples TS, OA provides an idea of how many sites are correctly classi ed (eq. 1), and spans from 0 (to corrected samples) to 1 (very accurate classi cation). 1 The Kappa coe cient K (eq. 2) is generated from a statistical test, and describes the accuracy of a classi cation compared to a random classi cation (Rwanga and Ndambuki, 2017;Aliani et al., 2019). Its value varies between 0 and 1, where 0 indicates a total accidental classi cation, while 1 indicates a very accurate classi cation. According to Gidey et al. (2017), good classi cations have K>0.8, while bad classi cations have K<0.4.
2 where the matric columns indicate the correspondence between ground truth data and pixel location, while the matrix rows indicate to which class is the pixel assigned.

Prediction of future LULC and associated driving forces
To manage natural resources (biodiversity) in uences, and to analyze and forecast spatial LULC changes, the Land Change Modeler (LCM) in TerrSet (formerly known as IDRISI) software was developed (Kumar et al. 2015;Aryaguna and Saputra 2020). LCM is an ArcGIS-integrated suite of tools for the assessment of future LULC changes, detecting gains and losses, net change, persistence and identi cation of transitions between LULC classes (Mishra et al. 2018). To map future LULC scenarios, LCM utilizes historical LULC maps and a series of driving forces (Table 3). The Markov chain projection is performed by creating matrixes to estimate the transition probability and the area of each LULC class for future dates (Hasan et al. 2020;Khoshnood Motlagh et al. 2021). In this study, LCM was applied to forecast the future LULC in three scenarios (2030,2040,2050), via a few main steps: i) analysis of historical LULC maps (1989,2004,2019) and associated changes, ii) creation of transition probability matrixes, iii) model validation, iv) prediction of future LULC maps, accounted for possible driving forces.
To  Leta et al. 2021a). Kappa values vary between 0 and 1, with values >0.8 meaning an almost perfect agreement. In detail, Kstandards is an index of agreement that attempts to account for the expected agreement due to random spatial reallocation of the categories in the comparison map; Kno is identical to Kstandards if both the quantity and allocation of categories in the comparison map are selected randomly; Klocation represents the extent to which the maps agree in terms of location of each LULC category.
To corroborate the study outcomes, a series of statistics were considered (Wang et al. 2016): agreement due to chance (agreement chance), agreement due to quantity (agreement quantity), agreement due to the location at the grid cell level (agreement grid cell), disagreement due to the location at the grid cell level (disagree grid cell) and disagreement due to quantity (disagree quantity) were calculated to indicate how well the comparison map agrees with the reference map (Wang et al. 2016).
Driving forces are the factors that affect LULC changes at the local scale, and therefore they should be locally investigated and addressed (Eastman 2009;Khan et al. 2021). In simulating future LULC, LCM differentiates between static and dynamic variables, where the rst are stable in time while the latter are temporally changing, and therefore recalculate every time-step (Table 3).
The Cramer's V Coe cient (CVC), sometimes also called Cramer's V strategies, was used to assess the correlations between the various driving variables. According to Eastman (2009), variables that have a Cramer's V>0.40 are good and these drivers will have the greatest impact on the modi cation process and its spatial distribution (Gaur et al. 2020;Benavidez-Silva et al. 2021;Vu et al. 2021

Historical LULC maps
Three reference years (1989,2004,2019) were considered to evaluate historical LULC, via a maximum likelihood supervised classi cation ( Figure 2). As reported in Table 4, in 1989 most of the study area was covered by agriculture (33%), grass/swamps (24%), and shrub (22%), with only a very minor part occupied by built-up (0.3%). Similar LULC was also observed in 2014, with agriculture (34%), grass/swamp (24%) and shrub (18%) being the most dominant LULC classes, and just a small increase in the area covered by built-up (1%). In 2019, the class distribution remained more or less similar, with an increase in the built-up area (1.75%). In summary, in the past, agriculture was always the most dominant LULC class in the Fincha watershed, followed by grass/swamp and shrub. The results reported in Figure

Accuracy assessment for historical LULC
The overall accuracies OA and Kappa values K were 82.80%, 85.57%, 89.82% and 80.51%, 82.54% and 87.84%, respectively, for the three reference years (Table 5 and supplementary Tables S1, S2, S3). These results indicate that the accuracy of the classi cations improved from 1989 to 2019, also thanks to the higher quality of the satellite data used. The accuracy of a map could be different for users and map developers. The user's accuracy indicates how often a speci ed class on the map is present on the ground, while the producer's (mapmaker) accuracy shows how frequently are real futures on the ground appropriately shown on the classi ed map or the probability that a certain land cover is classi ed according to eld evidence. Hailu et al. (2020) de ned the Kappa statistics <40%, 40%-75% and >75% as poor, good and excellent, respectively. Using this approach, from Table 5 one can notice that the statistics of the Fincha watershed were excellent, meaning a very good agreement between the classi cation maps and the reference information.

Historical LULC changes and transition probability matrix
Comparing the three reference years, it is possible to observe a considerable reduction in the area covered by forest and shrubs during the observation period (Table 6). In detail, yearly, around 639 ha of forest and 280 ha of shrub were cleared in favour of other LULC classes. As anticipated, human pressure contributed to changing the environment, as recognizable by the increase in areas covered by agricultural elds, built-up, grass/swamp and waterbodies, which yearly gained around 616 ha, 125 ha, 7 ha and 171 ha, respectively. Waterbodies increased signi cantly during the last 30 years, mainly because of human intervention. In fact, in 1989, the Amerti reservoir, one of the reservoirs located in the Fincha watershed, was not fully lled, while it was lled in 2014. In 2019, another dam was constructed over the Nashe River (Leta et al. 2021a). The study pointed out small changes in terms of grass/swamps, at least in terms of net variation. In fact, as visible from Figure 3 and Table 6, the majority of the Fincha watershed was affected by variations in LULC that include this class. The results presented here are in line with the existing literature on LULC in the Fincha watershed (e.g., Dibaba et al. 2020;Leta et al. 2021b). All the authors agreed that the shifting from natural LULC towards more anthropized environments could threaten biodiversity and decrease the total values of ecosystem services (Tolessa et al. 2021) The probability transitional matrix is the transfer direction of Land use land cover types from one category to other categories in the given year (Han et al. 2015). The nature change can be distinguished from the Markov transitional matrices for historical LULCC over the period between (1989-2004 and 2004-2019). The nature of change can be distinguished from the trend as depicted from the Markov transition matrices over the period between 1989 and 2019.
The diagonal values represent the probability that each land cover class remains persistent (constant) from earlier to later years. The other values represent a given land cover land class undertakes transition to another land cover land class.
Between 1989 and 2004, the highest and the lowest persistent LULC classes were waterbody and grass/swamps, characterized by a percentage of stability of 91% and 42%, respectively. During the period 2004-2019, the most and the less stable LULC class categories were waterbody and grass/swamp, which accounted for around 80% and 31%, respectively. Over the entire temporal horizon observed (1989 to 2019), a large part of the forest was converted to agriculture and grass/swamps (see Tables S4, S5 and S6 for the detailed LULC transition matrixes).
According to information obtained during eld investigations (KII and eld evidence), waterbodies increased after the newly constructed Nashe Dam. During the construction of the reservoir, many farmers along and downstream of the Nashe stream had been displaced to other agricultural places or towns. The abandonment of elds and the need for resettling in other areas caused a decrease in forest and an increase in built-up areas. In addition, not well-planned and long-term urban development and agricultural management strategies contributing to negatively affecting natural resources, causing its signi cant decline in the last decades (Figure 3).

Model validation
The LULC map of 2019, predicted from the 1989 and 2004 data, has been validated with the classi ed LULC map of the very same year (Table 7), showing that the LCM model can effectively forecast LULC changes. The capability of LCM in predicting the 2019 LULC was assessed via K-indexes and other statistics in TerrSet (Table 8).
All the values of k-indexes (>80%) indicate good agreement between the projected and the actual LULC map (Leta et al. 2021a). The Disagree Quantity (0.0742) is greater than the Disagree Gridcell (0.0268), indicating that the model has a higher ability to predict the LULC in location (spatial) than in quantity for the Fincha watershed. To corroborate the results presented here and to reduce the uncertainties, additional data should be included in the study, mostly deriving from laborious and expensive eld investigations. To forecast future LULC changes, it is needed to account for the most important driving variables (Table 9). As visible, all variables but the slope should be included in LCM. The LULC maps for 2030, 2040 and 2050 were created via LCM, using the historical maps as a basis ( Figure 4). As observed in the past, also for the future an increase in areas covered by agriculture, built-up, grass/swamp and waterbody is forecasted, while a drastic decrease in forest and shrub should be expected, with a slower rate of deforestation in the decade 2040-2050 ( Figure 5 and Table 10). Since both forest and shrub will continuously decrease in the future, a major effort in promoting protection measures to keep natural resources and biodiversity is advisable. In terms of transition probability (Tables S7, S8, S9), areas covered by forest and shrubs are more prone to be converted into agricultural land, while built-up areas should be expected on the actual grass/swamp zones. This indicates that, for the future, agriculture and built-up zones will expand at a high rate since the other LULC classes will be converted to them. Reversely, forests and shrubs will decline at a signi cant rate.
The LCM results point out that, in the incoming decades, signi cant changes in LULC should be expected, mostly because of an ever-increasing pressure of humans in need of more land for settlements and cropland. Indeed, the local population is growing, and more natural resources are needed to satisfy their request for food, energy and construction material (Leta et al. 2021a;Kenea et al. 2021). Besides direct consequences on the environment, the ongoing deforestation in the Fincha basin is also causing an intensi cation of soil erosion, triggering sediment siltation in the various reservoirs located in the region, eventually leading to a decrease in reservoir e ciency in terms of water availability and hydropower production.

Conclusions
The present study investigated the historical LULC (years 1989, 204, 219) in the Ethiopian Fincha watershed via a combination of satellite imagery and eld support data. Based on such analysis, the Land Change Modeller was applied to forecast LULC in the next three decades (years 2030, 2040, 2050). The 2019 LULC was also used for validating the LCM approach, comparing the forecasted situation with the actual one, indicating that the multi-layer Perceptron (MLP) neural network of Markov-Chain (MC) has enough capability to predict future LULC.
Over the last thirty years, the forest covering the Fincha watershed was mostly converted to agricultural and Grass/swamp areas. An increase in areas covered by waterbody and built-up was also observable, mainly connected with increasing human pressure and the construction of new hydropower reservoirs.
For the future, a similar trend is more than probable. Indeed, if management strategies will not change towards more sustainable ones, an even more signi cant decrease of the forest should be expected, in favour of new settlement areas and cropland. This change could help locals in sustaining their livelihood in the short term, but, in the medium/long term, the reduction of areas covered by forest will contribute to decreasing biodiversity and ecosystems services, as well as in fostering soil erosion, with detrimental consequences such as reservoir siltation.    Predicted LULC of the Fincha watershed for the next three decades.