Spatial Evolution of Prosopis Invasion and its E ﬀ ects on LULC and Livelihoods in Baringo, Kenya

: Woody alien plant species have been deliberately introduced globally in many arid and semi-arid regions, as they can provide services and goods to the rural poor. However, some of these alien trees and shrubs have become invasive over time, with important impacts on biodiversity, ecosystem services, and human well-being. Prosopis was introduced in Baringo County, Kenya, in the 1980s, but since then, it has spread rapidly from the original plantations to new areas. To assess land-use and land-cover (LULC) changes and dynamics in Baringo, we used a combination of dry and wet season Landsat satellite data acquired over a seven-year time interval between 1988–2016, and performed a supervised Random Forest classiﬁcation. For each time interval, we calculated the extent of Prosopis invasion, rates of spread, gains and losses of speciﬁc LULC classes, and the relative importance of Prosopis invasion on LULC changes. The overall accuracy and kappa coe ﬃ cients of the LULC classiﬁcations ranged between 98.1–98.5% and 0.93–0.96, respectively. We found that Prosopis coverage increased from 882 ha in 1988 to 18,792 ha in 2016. The highest negative changes in LULC classes were found for grasslands ( − 6252 ha; − 86%), irrigated cropland ( − 849 ha; − 57%), Vachellia tortilis -dominated vegetation ( − 3602 ha; − 42%), and rainfed cropland ( − 1432 ha; − 37%). Prosopis invasion alone directly accounted for over 30% of these negative changes, suggesting that Prosopis invasion is a key driver of the observed LULC changes in Baringo County. Although the management of Prosopis by utilization has been promoted in Baringo for 10–15 years, the spread of Prosopis has not stopped or slowed down. This suggests that Prosopis management in Baringo and other invaded areas in East Africa needs to be based on a more integrated approach.


Introduction
Woody alien plant species have been deliberately introduced in many arid and semi-arid regions of the world, as they can provide services and goods to the rural poor. Many of these have become invasive and continue to invade new regions at an unprecedented rate, exerting strong impacts on changes in the same area. Besides assessing LULC changes, we investigated how an exotic woody plant invaded, and the role that it played in fueling further LULC changes in the area. In order to achieve these goals, we use a time-series of multispectral and bi-seasonal Landsat data in combination with other important auxiliary data, and apply a novel machine learning classifier. A rigorous validation procedure makes sure that the obtained classification results are of high accuracy and allow for an accurate multi-temporal change analysis.

Study Area
This study was conducted in Baringo County, located along the Great Rift Valley system in Kenya (large map, Figure 1). It is characterized by a unique combination of altitude, precipitation, soil, and vegetation. The lowland "Njemps flats" are slightly undulating and dominated by rangeland, with an average altitude of 700 m above sea level. They are surrounded by Laikipia Escarpment on the east and Tugen Hills and Elgeyo Escarpment on the west, as well as the ridges and plateaus of the Lake Baringo catchment with peaks of over 3000 m a.s.l. [31]. The Tugen Hills form a conspicuous topographic feature which trends north-south and mainly consists of volcanic rocks. The steep slopes on escarpments end down the valleys forming prominent gullies and very deep gorges. The total annual precipitation is between 1000-1500 mm in the highlands [31] and between 300-700 mm in the lowlands, and is characterized by a bimodal distribution with two peaks in April and November [32]. Temperatures range from 10 • C to 35 • C. Presently, the vegetation is predominantly a woody mixture of indigenous and exotic species. It ranges from Vachellia-dominated deciduous shrubland on the valley floor to the evergreen forests in the highlands [33]. Prosopis juliflora currently dominates the lowland flats around the mid-east and mid-west of Lake Baringo and further stretching southwards up the northern tip of Lake Bogoria [10,21,31,33]. Historically, the flat parts in the lowlands of Baringo County consisted of a mosaic of grasslands and Vachellia-dominated shrubland [34,35].
Prosopis was introduced at these lowland zones between 1982-1983 through the Fuelwood Afforestation Extension Project [36,37]. The major objective of the project was to involve local communities in tree planting for mitigating problems such as lack of firewood and desertification [36]. In total, more than 20 plantations were established, covering an area of over 250 ha [10].
Our study area (red polygon area, Figure 1) focuses on the Marigat subregion because a) it is the zone where Prosopis was first introduced and invasion progress can be observed, b) Baringo County has three distinct eco-regions, but Prosopis is predominantly found in the lowlands, and c) because the most important LULC classes for local communities' livelihoods are concentrated in this zone, while the surrounding hilly and rocky escarpments are predominantly arid and have limited ecosystem services to offer. To analyze the relevant LULC changes, we focused on the LULC classes that are present in the invasion hotspot region around Lake Baringo and toward Lake Bogoria.

Field Reference Data
Field reference data were collected in the field between October 2015 and June 2016, which is a period that cuts across the dry and wet seasons. Reference points distributed across the study area were collected using a handheld GPS receiver. For the historical imagery, reference data was collected by the visual interpretation of VHR satellite imagery available via Google Earth Pro [38]-details are provided in supplementary material SI2-by visual interpretation of the Landsat satellite data itself in combination with field verification from local long-term residents familiar with the study area and its historic LULC. The selection of reference data was also highly informed by the distinct species occurrence characteristics such as along rivers, roads, vegetation differentiation because of elevation and soil types, homogeneity, and contiguity. Being a semi-arid zone, the vegetation is scanty; this is largely associated with historical degradation in the study area [30], which is a factor that enhanced identification. Additionally, the dominant vegetation is the generic species of Vachellia and Prosopis. Their identification was enhanced by seasonality as described in Section 2.2.2. For historical dates where VHR was not available, we heavily relied on local expert knowledge and vegetation occurrence characteristics to collect reference data. After the visual interpretation, we digitized small polygons, i.e. regions of interest (ROIs) for each LULC class (supplementary material SI2). We performed a classification on 13 LULC classes that were later aggregated to eight classes. 'Mixed vegetation', for example, is an aggregation of several landscape

Field Reference Data
Field reference data were collected in the field between October 2015 and June 2016, which is a period that cuts across the dry and wet seasons. Reference points distributed across the study area were collected using a handheld GPS receiver. For the historical imagery, reference data was collected by the visual interpretation of VHR satellite imagery available via Google Earth Pro [38]-details are provided in Supplementary Materials SI2-by visual interpretation of the Landsat satellite data itself in combination with field verification from local long-term residents familiar with the study area and its historic LULC. The selection of reference data was also highly informed by the distinct species occurrence characteristics such as along rivers, roads, vegetation differentiation because of elevation and soil types, homogeneity, and contiguity. Being a semi-arid zone, the vegetation is scanty; this is largely associated with historical degradation in the study area [30], which is a factor that enhanced identification. Additionally, the dominant vegetation is the generic species of Vachellia and Prosopis. Their identification was enhanced by seasonality as described in Section 2.2.2. For historical dates where VHR was not available, we heavily relied on local expert knowledge and vegetation occurrence characteristics to collect reference data. After the visual interpretation, we digitized small polygons, i.e., regions of interest (ROIs) for each LULC class (Supplementary Materials SI2). We performed a classification on 13 LULC classes that were later aggregated to eight classes. 'Mixed vegetation', for example, is an aggregation of several landscape components (i.e., sub-classes). The criteria for aggregation were informed by our study objectives and the LULC present in the invaded areas. The defined classes are presented and described in Table 1. We did not include settlements because the area is sparsely populated, and settlements are pastoral in nature; they are small semi-permanent structures that are generally not easily identifiable on 30m Landsat data. This study focuses on a period between 1988-2016, for which Landsat satellite data with a spatial resolution of 30m were available. We chose cloud-free good quality imagery of 1988 (base year) because before this date, Prosopis was barely visible on Landsat data, and invaded areas were small and mostly within the plantations' perimeter.
We downloaded Landsat data from the Landsat Surface Reflectance Climate Data Record (Landsat CDR), which has been atmospherically corrected [39,40]. For all Landsat datasets, we selected the bands blue, green, red, near infrared (nir), as well as the two shortwave infrared bands (swir1 and swir2). In consideration of the limitation posed by multispectral sensors such as Landsat in mapping vegetation species, we enhanced the ability of our data to detect the species in our study area by additionally calculating the normalized difference vegetation index (NDVI) for both the dry and wet season. NDVI is a ratio-based index, which introduces information about biophysical properties, thereby increasing the separability of landscape components [41]. We considered both the dry and wet season because our study context is a semi-arid zone with two distinct seasons. The choice of the two periods allowed the maximization of the spectral differences between Prosopis and other native vegetation such as the dominant Vachellia spp. and seasonal herbaceous vegetation. Prosopis is evergreen, unlike the deciduous Vachellia; hence, it tends to maintain a higher vigor and canopy than the native vegetation during the dry season [26]. Per temporal step, satellite images were selected between January and March for the dry season, and between April and August for the wet season ( Table 2). To further improve our data in each combined dataset, we also included the digital elevation model generated by the Shuttle Radar Topography Mission (SRTM). In the study region, Prosopis grows exclusively at elevations below 1500 m a.s.l [42]. In total, we created five data stacks consisting of 15 bands each. Pre-processing included mosaicking, cloud masking, and the stacking of dry and wet season bands. To ensure optimal data quality, we manually selected the best images; the main criteria were (i) the acquisitions date, and (ii) the absence of cloud cover. For classification, we used the Random Forest (RF) algorithm, which we implemented using the function provided in the randomForest package in R statistical software version 3.4.4 [43,44]. RF is a non-parametric machine learning algorithm that consists of an ensemble of randomly grown trees, whereby predictions are made by each individual tree, which are then averaged [45]. Each tree is grown using a different random bootstrap sample with two-thirds used for building the random forest model, and one-third used for estimating the prediction and the associated error, which is also known as the 'out-of-bag' (OOB) error [46][47][48]. Once the RF model is built using reliable reference samples for training, the RF model can then be applied to the data to be classified. In this study, we generated a model for each point in time. We used 70% of the manually delineated ROIs described in Section 2.2.1 for training each RF model, and then applied it to the corresponding satellite data stack. Nearly one-third (30%) of training samples were used for validation purposes. For each RF model, we used 1000 trees. All the other arguments of the randomForest function were kept at the default setting [48].

Accuracy Assessment
Besides the calculation of the OOB, we performed an additional and independent 10-fold cross-validation accuracy assessment. Cross-validation (CV) is a robust statistical method that is used to evaluate the model accuracy by dividing the reference data into a training and a validation set. It assumes that data are identically distributed and that training and validation samples are independent [47]. In k-fold cross-validation, data are first randomly partitioned into k equally (or nearly equally) sized folds. Subsequently, k iterations of training and validation are performed such that within each iteration, a different fold of the data is used for validation, while the remaining k-1 folds are used for learning [49]. The validation is performed on a pixel level, so that each pixel inside a validation polygon is compared with the reference class. Then, the k results from the folds can be combined to produce a single estimation. In this study, we ran a 10-fold cross-validation, where the training dataset was randomly split 10 times with 90% for training and 10% for validation, respectively. Then, the 10 results for each classification were summarized, and confusion matrices were generated for the calculation of overall, users', and producers' accuracies (OA, UA, PA) as well as Cohen's kappa (K). Cross-validation was performed independently for each of the five classification outputs.

Land-Use and Land-Cover Change Analysis
We assessed LULC change by creating cross-tabulation matrices for the intervals from 1988 to 1995, 1995 to 2002, 2002 to 2009, 2009 to 2016, and 1988 to 2016, and subsequently calculated gains, losses, net changes, and rates of change. LULC gains and losses were visualized using charts and tables. In addition, interval-specific Prosopis areal coverage and invasion rates were determined. To determine how Prosopis invasion impacted individual LULC and at which time interval, the net changes per class per time interval were derived from subtracting the contributions of each class to Prosopis (Losses to P) from their respective gains from Prosopis (Gains from P).
To understand the invasion pattern originating from the initial plantations, we analyzed change statistics and visually assessed the LULC maps. Knowledge of seasonal socio-economic activities such as irrigation farming, charcoal production, and flood events also informed the interpretation of the change output and distribution patterns. This knowledge was acquired through field observations, focus group discussions, and interactions with the host community.

Results
Random Forest classifier achieved acceptable to good producer's and user's accuracies of above 70% for all classes, except for V. tortilis (Forssk) Galasso and Banfi in 2016 (PA = 53.5%) and for Prosopis in 2002 (PA = 64.8%). Overall accuracy ranged between 98.1-98.5%. The kappa coefficient ranged between 0.93-0.96. Generally, the highest accuracies were realized for the 2016 classification, and the lowest accuracies were realized for the 1988 classification (Tables S1-S5: Supplementary Materials).

Invasion History of Prosopis from 1988 to 2016
In 1988, the dominant LULCs in the 180,000-ha large study area were mixed vegetation and bareland at 71.5% and 8.3% of total study area, respectively ( Figure 2). At that time, there was subtle Prosopis coverage mainly on the shores of Lake Baringo and a few Prosopis patches in the Eldume and Kailer sublocations. Initially, the prominent invasion pathways were mostly along the rivers (i.e., Molo, Perkerra, Loboi, Endao, Lembus, Chemeron, and Sandai), road networks, and irrigation canals ( Figure 2B). In Figure 2C, an invasion corridor from Lake Baringo southwards to Lake Bogoria through the water marshes (that currently form Lake 94, which is located between lakes Baringo and Bogoria) and the riparian zones of rivers draining into the lakes can be identified. From this corridor, Prosopis continued to spread into the surrounding areas, covering largely bareland predominantly on the south and southwestern zones of Lake Baringo. This led to the invasion of the Perkerra irrigation scheme (indicated in Figure 2A), which is a key agricultural facility that was established during the colonial times in the 1950s. In contrast to the rapid spread of Prosopis in the southeastern part of the study area, its spread toward the northern and eastern zones of Lake Baringo appears to be considerably slower.
Over the study period, there was a sharp and steady increase in Prosopis coverage, from 882 ha in 1988 to 18,792 ha in 2016 (Table 3, Figure 3A,B). However, Prosopis did not spread at the same rate over that time (Figure 3). The rate in the seven-year intervals increased continuously from 4.7% until 1988 to 13.1% in 1995 and stabilized at around 27% since 2002. This can be translated into a yearly spread rate of 3.8-4.0% per year since 2002. Over the study period, there was a sharp and steady increase in Prosopis coverage, from 882 ha in 1988 to 18,792 ha in 2016 (Table 3, Figure 3A and B). However, Prosopis did not spread at the same rate over that time (Figure 3). The rate in the seven-year intervals increased continuously from 4.7% until 1988 to 13.1% in 1995 and stabilized at around 27% since 2002. This can be translated into a yearly spread rate of 3.8-4.0% per year since 2002.

Land-Use and Land-Cover Changes between 1988 and 2016
During the 28-year period, two of the eight LULC classes (Prosopis and water) increased in spatial coverage by 2031% and 42%, respectively, while the rest decreased (Table 4 and Figure 3A). Increases in the Prosopis invaded area is more or less constant since 1995, and was not affected by the implementation of 'management by utilization' and the flooding event in 2013 ( Figure 3A,B). Water levels first declined, but started to increase again since 2009 ( Figure 3A). Overall, the highest losses were found in grasslands by 86%, followed by irrigated cropland at 57%, bareland at 43%, V. tortilis at 42%, rainfed cropland at 37%, and the least was mixed vegetation at 4% (Table 4). However, some the classes that generally declined also showed some spontaneous slight increases during certain intervals such as bareland in 1995, croplands in 2002 and 2009, and V. tortilis in 2016 (Figure 4). Prosopis showed a steady increase across all the intervals (Table 4 and Figures 3A  and 4). In the early years (1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)

Prosopis-Specific Induced Changes on other LULC 23
Our results show that V. tortilis and grasslands were mostly lost to Prosopis during the period 24 1995 to 2002 (by 22% and 12% respectively), while mixed vegetation, bareland, and irrigated 25 cropland were mostly lost to Prosopis during the period 2009 to 2016 (by 1.94%, 18%, and 74%, 26 respectively; Table 5 and Figure 5). Generally, Prosopis induced the most changes in LULC in the 27 study area between 2009-2016, while the least impacts were observed between 1988-1995. 28 Cumulatively, within almost three decades, Prosopis invasion has resulted in LULC losses of V. 29 tortilis by 3452 ha (41%), grasslands by 2675 ha (37%), bareland by 5351 ha (36%), irrigated cropland 30 by 373 ha (25%), mixed vegetation by 6215 ha (5%), and rainfed cropland by 129 ha (3%). The three 31 largest donors to Prosopis were, in order of ranking, mixed vegetation, which comprises mostly 32 indigenous trees and shrubs, bareland, and V. tortilis, with a mean annual loss of 222 ha/a, 191 ha/a, 33 and 123 ha/a, respectively. Grasslands and irrigated croplands, which are the basic livelihood 34 support system for the agro-pastoral community, annually lost 96 ha and 13 ha, respectively. 35

Prosopis-Specific Induced Changes on other LULC
Our results show that V. tortilis and grasslands were mostly lost to Prosopis during the period 1995 to 2002 (by 22% and 12% respectively), while mixed vegetation, bareland, and irrigated cropland were mostly lost to Prosopis during the period 2009 to 2016 (by 1.94%, 18%, and 74%, respectively; Table 5 and Figure 5). Generally, Prosopis induced the most changes in LULC in the study area between 2009-2016, while the least impacts were observed between 1988-1995.
Cumulatively, within almost three decades, Prosopis invasion has resulted in LULC losses of V. tortilis by 3452 ha (41%), grasslands by 2675 ha (37%), bareland by 5351 ha (36%), irrigated cropland by 373 ha (25%), mixed vegetation by 6215 ha (5%), and rainfed cropland by 129 ha (3%). The three largest donors to Prosopis were, in order of ranking, mixed vegetation, which comprises mostly indigenous trees and shrubs, bareland, and V. tortilis, with a mean annual loss of 222 ha/a, 191 ha/a, and 123 ha/a, respectively. Grasslands and irrigated croplands, which are the basic livelihood support system for the agro-pastoral community, annually lost 96 ha and 13 ha, respectively.

Spatial Evolution of Prosopis Invasion
Owing to its invasive nature, Prosopis has the potential to colonize and dominate a variety of landscapes, especially in the tropical ASALs [50]. However, despite this well-known fact, there remains a significant lack of understanding of the rate of invasion in affected regions and the consequences for LULC dynamics. We assessed the evolution of Prosopis invasion and the associated LULC changes over time and space by using multi-temporal and multispectral Landsat satellite imagery as well as by verification of mapping outputs with historical Google Earth Imagery [38] and local expert knowledge. Applying Random Forest on the Landsat data provided good estimations and reliable accuracies, which has also been reported in other studies [21,[51][52][53][54]. This provides potential for the long-term monitoring of Prosopis invasion. Moreover, its processing chain can be easily repeated, which is particularly useful when performing time-series analysis, and can also be applied to other studies with some fine-tuning [21]. Our study covered the entire invasion history, and used rigorous methods to assess the evolution of Prosopis. We showed how the invasion dynamics have influenced LULC in the study area, which was information that was lacking in previous studies. The use of seasonality, especially the dry season information, was very useful in the accurate mapping of Prosopis from the coexisting native vegetation and other LULC. However, we would like to note that our study was limited by inaccessibility to the areas to the east of Lake Baringo. This is a hotspot zone for community-based violence [55] and at the time of field data collection, the area was quarantined. Therefore, we did not succeed in collecting ground truth information for that particular zone.
Zooming into a heavily invaded region with original Prosopis plantations further illustrates how the combination of multi-temporal Landsat satellite imagery and Random Forest helped in identifying and visualizing invasion processes at multiple scales ( Figure 6). Both the temporal and spatial patterns detected at the case study level and at smaller scales revealed that Prosopis initially spread away from plantations mainly along roads and waterbodies or lakeshores. This is in line with findings from the Afar region in Ethiopia, where waterbodies and roads have been identified as key drivers of Prosopis invasion [16]. Subsequently, it invaded areas away from waterbodies, particularly V. tortilis, grassland, and mixed vegetation and more recently also, mixed vegetation on the slopes of the surrounding hills, particularly by following riverbeds upstream. This raises the question of to what extent and at what rate Prosopis will also be able to displace mixed vegetation on the surrounding hills and escarpments, which was facilitated by Prosopis seed dispersal by migrating livestock [16].   It is likely that in our analysis, grasslands were partially under-represented or over-represented at certain intervals. This is because some grasslands in the study area are dominated by annuals, while the perennial grasses are stressed due to overgrazing. As such, grasslands are more abundant during the wet season when weather conditions are favorable, which is a time when other herbaceous vegetation is also at its peak. Furthermore, it was challenging to separate dry perennial grasses from bareland. Another challenge in our analysis was the differentiation between V. tortilis and Prosopis along rivers and in wetlands. It was relatively easy to differentiate V. tortilis from Prosopis during the dry season in areas away from water, because, unlike Prosopis, it sheds leaves. However, in riparian habitats, V. tortilis, as well as another common native, Balanites aegyptica, sometimes behaves similar to evergreen species, due to the close proximity to water. All these species are closely related in spectral reflectance [21], which posed a challenge in their effective discrimination in the riparian and wetland zones using Landsat data.

Spatial Changes in Prosopis Coverage
In 1982 and 1983, Prosopis was introduced in 31 locations in the study area [10] to alleviate fuelwood scarcity and to rehabilitate heavily overgrazed drylands [13,36,50]. These plantations were established within the productive lowland zones along rivers, grasslands, and irrigation schemes in order to enhance the tree survival rate in this harsh semi-arid environment. These zones are more prone to invasion than other areas, because they are convergent zones for waterborne and animal-dispersed seeds, and are more conducive for plant growth [22]. Our findings that by 1988 Prosopis covered some 882 ha is comparable to estimates reported by Mwangi and Swallow [33] (739.5 ha by 1990). Our estimate of the average annual increase in Prosopis coverage (640 ha) is within the range reported by Kyuma et al. [25], who reported that Prosopis spread at an average annual rate of around 532 ha/a on hillslopes and of 13 to 1309 ha/a in the floodplains in Kajiado County in southern Kenya. Visual comparisons of our Prosopis map with the mapping output from Ng et al. [21] using higher spatial resolution data (Sentinel-2 and Pléiades) also show an agreement of the invasion distribution. Hence, our results confirm that Prosopis is an aggressive invader in Baringo County, which is in line with observations from other regions in Africa [16,27,[56][57][58][59]. The rapid spread of the species can be explained by different factors: (a) the massive reproduction of seeds, (b) effective seed dispersal agents (e.g., livestock and wildlife) and pathways (rivers and roads), or (c) favorable climatic conditions. The flooding event in 2013 [21] may also have contributed to the further spread of Prosopis. Figure 6 reveals a discriminative invasion pattern from the plantations. As already observed by Andersson [10], the establishment of an initial plantation did not automatically result in the invasion of the surrounding habitats. However, the areas (Sandai, Logumgum and Loruk) that initially remained free from Prosopis invasion became invaded eventually, i.e., about a decade after these plantations were established ( Figure 6). It should be noted that some sites received P. juliflora only, some received P. pallida only, and some sites received both species [13,60]. Prosopis juliflora is known to be highly invasive, while Prosopis pallida appears to be less invasive, or not invasive at all in Baringo County (Purity Rima Mbaabu, personal observations). It remains to be shown whether the varying level of invasion around plantations is driven by the taxonomic composition of the plantations or by variation in the local environmental conditions [10], which may lead to a 'time lag' in the invasion process. Our findings suggest that Prosopis invasion in our study area follows a regular and gradual pattern of increase.

LULC Changes in the Study Area
The Lake Baringo ecosystem presents a good example of a semi-arid landscape experiencing remarkable LULC changes over the course of a few decades [30]. While the area invaded by the alien Prosopis increased more than 20-fold since 1988, LULC classes with native vegetation and agricultural areas decreased. Thus, the local communities have been losing land-cover classes that form part of their key livelihood support system, such as grasslands and irrigated croplands, at an alarming rate. Our results corroborate the findings of Kiage et al. [30], who assessed LULC changes around Lake Baringo also using Landsat imagery, and found that the forest area (which corresponds to our V. tortilis class) dropped by 40% within a 14-year period (1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000). Our study shows that Prosopis invasion directly accounts for over 30% of the LULC changes in the study area. We would like to note that this may be an underestimation of the current state of invasion, because, a) considering that there is usually a lag-phase between Prosopis introduction and the actual detection on satellite image, and b) the use of a 30 m Landsat imagery poses a challenge in detecting young but sparse Prosopis invasions.
Besides Prosopis invasion, human activities such as deforestation, land clearing, and overgrazing, together with climate change, are likely to have contributed to the observed LULC dynamics in Baringo County [30,61]. Yet, regardless of the factors triggering the LULC dynamics, Prosopis tends to be more adaptive to the observed environmental change than native plant species [62,63]. This is because, due to its deep-reaching roots, Prosopis is able to survive in areas where nothings else seems to grow [14,64]. In addition, Prosopis has been shown to directly interfere with the survival of native plant species, probably via competition for water or allelopathy [65]. Losses in key agro-pastoral livelihood supporting LULCs (i.e., V. tortilis, grassland, and irrigated cropland) showed variable patterns across the study period; while the decline in areas covered by V. tortilis and grasslands was more prevalent during the first decade after Prosopis introduction (1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002), the highest losses in rainfed and irrigated cropland occurred between 2009-2016 ( Table 3). Most of the V. tortilis and grassland areas still present at the beginning of the study period were at close proximity to the plantations (Figure 6), which may explain why they were first invaded by Prosopis.
Over the last 100 years, Baringo County has experienced a history of land degradation [10,61], and the increase in bareland between 1988-1995 may be considered as a continuation of this trend (Table 3). This matches with the findings of Kiage et al. [30], who observed that bare areas increased around Lake Baringo area between 1986-2000. Land degradation in Baringo County was likely to be caused by a combination of an increase in human population and livestock numbers, as pastoralism used to be the main economic activity for local communities [66]. Additionally, low levels of precipitation experienced in the area between 1978-1996 (annual average = 51.7 mm) may have further contributed to the increase of bareland [67]. Variation in rainfall is particularly pronounced in semi-arid areas such as the Baringo lowlands. Moreover, the long-term average precipitation in the region shows a slight downward trend since the 1960s [68]. This may also explain why the lakes' water levels declined between 1988-2002.
However, since 2009, the area under water increased again, which was a phenomenon that has also been observed in other waterbodies within the Great Rift Valley system. While the causes of this increase remain unclear to date, some sources attribute the increase to increased rainfall, to plate tectonic movements, or to land-use changes and its associated siltation downstream [69].
Agricultural activities started to increase in Baringo County at the beginning of the 21st century, but the cultivated area considerably decreased again after 2009 (Tables 3 and 4). The initial increase in irrigated cropland resulted from an expansion of irrigation schemes in the study area [70]. Small-scale irrigated farming activities in the area are largely seasonal in nature. They depend on water availability and seed production demand by the contracting seed companies such as Kenya Seed [71]. These companies contract farmers for various seed production, which is usually on a rotational and annual basis. However, if in a certain year farmers are not engaged in this activity, their farms usually remain idle, which leaves the farms vulnerable to Prosopis invasion, particularly after crop harvesting, when the livestock keepers freely graze their animals, which is a key vector of Prosopis seed dispersal, on crop residues [72]. For instance, in 1995, there were minimal farming activities in the Perkerra irrigation scheme. This coincided with the start of Prosopis invasion in the irrigation scheme, which eventually resulted into rapid Prosopis spread in subsequent years (Figure 7). It should be noted that most of the irrigated and rainfed cropland within the boundaries of the area invaded by Prosopis actually also bear a significant Prosopis soil seed bank, and will thus need increased maintenance efforts; otherwise, this land will be reinvaded within a few months (Purity Rima Mbaabu, personal observations). Finally, the drylands of Baringo are often characterized by inter-community violence [55]. The main cause of these conflicts is competition for scarce but vital resources, particularly water and fodder for livestock. Invasion into the grasslands has significantly reduced available fodder, because high densities of Prosopis (>50% canopy cover) have been reported to suppress grass growth and In the study area, charcoal production is a thriving economic activity with charcoal from V. tortilis being preferred on the market to that from Prosopis [21]. As such, V. tortilis was intensively harvested for charcoal production in the years before the millennium. Around 2007, during a National Field Day, the Kenya Forest Service (KFS) in Marigat expressed the need to ban the utilization of V. tortilis [73]. As an alternative, the communities were encouraged to utilize Prosopis, which was also regarded as a way of managing Prosopis invasion. The observed increase in V. tortilis after 2009 might be an effect of this ban, but we would like to point out that our accuracy assessment suggests a slight overestimation of V. tortilis in the southeast of our study area in the 2016 image. This arose due to misclassifications of this particular class, as depicted by the low producers' accuracy of 53.5% (Supplementary Materials;  Table S5), which may partly be explained by our study limitation already described in paragraph 1, Section 4.1.
By 2005, Prosopis invasion had intensified, and was perceived by the pastoralists as the cause of the dwindling grasslands for their livestock, which prompted a legal suit between the host pastoral community and the Kenyan government [74]. In response, the government began a sensitization program through which the affected communities were trained on Prosopis management by manual or physical removal and reseeding the cleared areas with grass species [13,29,75]. This practice was gradually adopted by the pastoralists as a new way of enhancing the fodder that was available for their livestock and income generation through commercial grass seed production. It has consequently led to an increase in the number of grass fodder farms in the area [75], and might explain the increase of grasslands in the last observed interval, between 2009-2016. However, our results show that the adoption of the 'management by utilization' approach promoted by the Kenyan government has not been effective in curtailing the further spread of Prosopis. On the contrary, utilization results in the coppicing of multiple Prosopis stems from the harvested stem, which creates impenetrable thickets and fuels the further spread of Prosopis through multiplied seed production by coppiced stems [59,76,77]. Building on the long-term management experience gained in Australia, we propose that sustainable and effective Prosopis management can only be achieved by combining different management options such as chemical, biological, and mechanical control. In addition, a regular inspection of areas at risk of invasion and the targeted removal of small Prosopis populations at the invasion front is recommended [77].
Another interesting observation was the expansion of agricultural activities, particularly irrigated cropland, which increased between 2002-2009. During this time, several seed companies contracted farmers to produce cereal seeds. Small-scale irrigated farming activities in the area are largely seasonal in nature. They depend on water availability and levels of seed demand by the contracting seed companies, which vary from one year to another. As such, if in a certain year farmers are not engaged in this activity, their farms usually remain idle. This inactivity leaves the farms vulnerable to Prosopis invasion, particularly after crop harvesting, when the livestock keepers freely graze their animals on crop residues, which is a key pathway for Prosopis seed dispersal in the croplands [72,78]. For instance, we observed that in 1995, there were minimal farming activities in the Perkerra irrigation scheme. This coincided with the start of Prosopis invasion in the irrigation scheme, which eventually resulted into rapid Prosopis spread in the subsequent years ( Figure 7). Our observation suggests that active land use plays a critical role in controlling Prosopis spread.
Finally, the drylands of Baringo are often characterized by inter-community violence [55]. The main cause of these conflicts is competition for scarce but vital resources, particularly water and fodder for livestock. Invasion into the grasslands has significantly reduced available fodder, because high densities of Prosopis (>50% canopy cover) have been reported to suppress grass growth and reduce understory species diversity [10,[79][80][81]. This is likely to exacerbate these resource-use conflicts between the local communities. Moreover, high-density Prosopis cover impedes accessibility to water points for people and livestock, and limits livestock mobility to grazing areas [50]. Further, Prosopis is known to consume substantial amounts of groundwater [82][83][84][85], but the effect of Prosopis invasion on the groundwater level in Baringo County is currently unknown.

Conclusions
Baringo County has a long history of LULC changes, even prior to Prosopis invasion. Historically, the main drivers of LULC changes are considered to be a growing human population and an increase in livestock numbers, which lead to an unsustainable exploitation of the natural resources in the region. This in turn prompted the introduction of Prosopis and other exotic tree species in Baringo County and other parts of Kenya.
Our study provides evidence that over the last 30 years, Prosopis has become an additional driver of LULC changes in Baringo County. While Prosopis has become a key source of wood and the invasion of bareland can improve the microclimate in the area and increase soil stability, the substantial invasion of grasslands, Vachellia tortilis-dominated land, and cropland as revealed in this study has serious consequences for the delivery of traditional ecosystem services provided to (agro-)pastoralists, and thus on the livelihood of rural people living in the invaded area. Based on our findings, we draw the following conclusions:

1.
Freely available Landsat data analyzed with the implementation of Random Forest machine learning algorithm in the open source R software are useful in assessing spatial temporal LULC changes, especially in regions where commercial data and software is economically inaccessible.
The RF algorithm has the ability to separate various landscape components with reliable accuracies.

2.
The use of bi-seasonal (dry and wet) multispectral data combinations and the Random Forest algorithm allowed us to spatially and quantitatively investigate the evolution of Prosopis invasion, its current extent, and the changes in LULC in the semi-arid environment. The use of dry season images enhanced the ability to differentiate the evergreen Prosopis from deciduous native Vachellia species on the medium-resolution Landsat imagery.

3.
Prosopis has rapidly increased in the study area since its introduction in 1982. It currently stands at approximately 18,792 ha, invading at a rate of 640 ha per annum. This rapid spread was facilitated by effective seed dispersal agents such as livestock and wildlife, a lack of natural enemies, and favorable climate, among other factors. Our findings indicate that Prosopis has been a key driver of LULC changes in the semi-arid lowlands of Baringo, directly accounting for over a third of the LULC changes observed over the last three decades.

4.
The LULC classes most vulnerable to Prosopis invasion are grasslands, V. tortilis-dominated zones, and fallow irrigated agricultural fields. Their vulnerability is enhanced by their ecological niche also favoring Prosopis growth and the presence of Prosopis seed dispersal vectors such as livestock, wildlife, water, and humans.
Our study provides important baseline data and information for policy management. The results indicate that the 'management by utilization' approach that has been promoted in Baringo for 10-15 years has not curtailed the further spread of Prosopis. In order to achieve the sustainable management of Prosopis and reverse its adverse impacts manifested in the loss of ecosystem services and livelihood to the local communities, we suggest that this species be targeted with an integrated and coordinated management strategy. negative impact on ecosystem services and rural livelihood" (Grant Number: 400440_152085). Urs Schaffner was supported by CABI with core financial support from its member countries (see http://www.cabi.org/about-cabi/ who-we-work-with/key-donors/).