You are on page 1of 12

Land Use Policy 87 (2019) 104051

Contents lists available at ScienceDirect

Land Use Policy


journal homepage: www.elsevier.com/locate/landusepol

The effects of trainings in soil and water conservation on farming practices, T


livelihoods, and land-use intensity in the Ethiopian highlands

Nathan S. Chestermana,b, ,1, Julia Entwistlea, Matthew C. Chambersa, Hsiao-Chin Liua,
Arun Agrawala, Daniel G. Brownc
a
School for Environment and Sustainability, University of Michigan, Ann Arbor, MI, United States
b
Environmental Incentives, Washington, D.C., United States
c
School of Environmental and Forest Sciences, University of Washington, Seattle, WA, United States

A R T I C LE I N FO A B S T R A C T

Keywords: Smallholder farmers in the Ethiopian Highlands face increasingly difficult farming conditions. Agricultural in-
Soil conservation tensification to feed the growing rural population, livestock pressure on native vegetation, and climate change
Smallholder agriculture converge to exacerbate soil erosion, creating a significant threat to crop productivity and rural livelihoods.
Farmer training Farmer trainings can be a potent strategy to increase farmer awareness of the larger causes of soil loss, help
Land-use change
farmers adapt to or mitigate environmental challenges, and improve crop outputs and incomes. However,
Land-use land-cover classification
evaluations of farmer training programs rarely assess socioeconomic and environmental outcomes simulta-
Ethiopia
neously. Our study uses multiple methods to estimate the socio-ecological effects of a soil and water conservation
training program for farmers implemented by ADHENO Integrated Rural Development Association, an Ethiopian
non-governmental organization. We asked: (a) did farmers use practices taught in trainings; (b) did these
strategies lead to improved agricultural productivity or livelihoods; and (c) how did land-use intensity change in
areas with more participation in farmer trainings? To address questions (a) and (b), our study and results are
based on 449 household surveys, collected between June and August 2017, in two rural kebeles in Ethiopia in
which ADHENO has been active since 2003. We used remote sensing analysis of high resolution satellite images
to measure land-use change over a period of nine years, addressing question (c) for one of the kebeles. Analysis
of household surveys with propensity score matching and robustness checks suggests that participation in farmer
trainings predicts the implementation of four out of seven soil conservation farming methods. Participants in
training sessions also had higher average incomes from agriculture than non-participants. Linear models of land-
use change in the area around villages did not depict a clear relationship between participation in farmer
trainings and change in land-use intensity. These results indicate that farmer trainings focusing on soil and water
conservation have the potential to influence farming practices and livelihoods and have implications for soil
conservation efforts in the Ethiopian Highlands. This study highlights the importance of evaluating the impacts
of small, localized development interventions, of which there are many in sub-Saharan Africa, to better un-
derstand the ways in which myriad types of programs influence both land use and livelihoods.

1. Introduction intensification, and deforestation (Anley et al., 2007; Haregeweyn


et al., 2015; Holden and Shiferaw, 2004; Shiferaw and Holden, 1999).
Approximately 80% of Ethiopian households practice smallholder The relationship between these factors is complex and multi-
agriculture (World Bank Group, 2015). This proportion has remained directional. Population growth, persistent rural poverty, and decreasing
stable since 2000, even as the national poverty rate fell from 45% to soil productivity may all contribute to farmers’ decisions to intensify
30% over the same time period (World Bank Group, 2015). None- agricultural production, which further exacerbates underlying resource
theless, across Ethiopia, the soil and water resources that smallholder degradation issues (Delelegn et al., 2017; Lambin et al., 2001). Ad-
farmers depend on are at significant risk of degradation due to several ditionally, climate change has further exacerbated the situation, as in-
factors, including climate change, population growth, agricultural creasingly unpredictable weather patterns negatively affect farm


Corresponding author at: University of Michigan School for Environment and Sustainability, 440 Church St., Ann Arbor, MI, 48109, United States.
E-mail address: nchest@umich.edu (N.S. Chesterman).
1
Present Address: 725 15th St. NW, Washington, D.C. 20010.

https://doi.org/10.1016/j.landusepol.2019.104051
Received 21 September 2018; Received in revised form 9 June 2019; Accepted 11 June 2019
0264-8377/ © 2019 Elsevier Ltd. All rights reserved.
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

productivity (Deressa and Hassan, 2009; Holden and Shiferaw, 2004). kebeles. ADHENO’s mission is “improving the living conditions of the
As smallholder farm productivity plateaus or decreases, farmers are rural people of North Shewa by working with communities and closely
forced to choose between intensifying land-use to continue production collaborating with the government…through participatory designing,
– placing soil health and other ecosystem services at greater risk of implementation, and evaluation of projects and programs of environ-
degradation – or, reducing land-use intensity to preserve soil fertility mental rehabilitation” (ADHENO, 2015, p. 10). In fulfillment of its
for future planting seasons, even if it might mean a smaller harvest for mission, ADHENO works with government extension agents to engage
the time being (Adimassu et al., 2016; Delelegn et al., 2017; Lambin rural farmers in trainings on techniques intended to improve soil sta-
et al., 2001; Tolessa et al., 2017). bility, climate change adaptation, and agricultural livelihoods. The
In the rural highlands, where the rugged topography facilitates soil practices covered by these trainings include digging trenches, terracing,
erosion, and there is a high population density and a millennia-long planting vegetation to stabilize soil, use of improved seeds, intercrop-
history of agropastoralism (Anley et al., 2007; Baye, 2017; Darbyshire ping, fertilizer application, and irrigation (ADHENO, 2015). It should
et al., 2003; Nyssen et al., 2004), these issues converge to create food be noted that these training topics comprise soil conservation techni-
shortages and the potential for poverty trap dynamics, stoking a sense ques that are traditionally practiced in the Ethiopian highlands, or
of urgency in the Ethiopian government and among the global aid newer approaches commonly covered by agricultural interventions
community to address soil erosion and improve smallholder farm pro- (Haregeweyn et al., 2015; Kato et al., 2011). Trainings are held in
ductivity (Sachs et al., 2004; Shiferaw and Holden, 1998; Tekle, 1999). centrally-located villages throughout two kebeles, with up to 40
Though farmers have been practicing traditional soil and water farmers invited to each training from the surrounding villages. Train-
conservation techniques for hundreds of years (Osman and Sauerborn, ings last for up to five days and cover one or two of the aforementioned
2001), these methods were used sporadically until the Ethiopian gov- practices. Invitations are extended to farmers that ADHENO field staff
ernment institutionalized soil and water conservation programs in the assess are likely to utilize the methods based on the suitability of their
1970s to address resource degradation on a national scale (Haregeweyn farm plots for implementing the practices covered in trainings. Some
et al., 2015). The most common conservation techniques include dig- trainings include incentives for participation – notably, farmers are
ging trenches to capture runoff, establishing exclosures to rehabilitate provided improved seeds during trainings on that topic to facilitate
vegetation, and constructing dams across gullies (Haregeweyn et al., uptake. Additionally, during some trainings on farming practices that
2015). The government implements these strategies by coordinating involve landscape modification, such as terracing and digging trenches,
with aid partners to engage farmers in watershed-level public works farmers are given short-term access to hand tools. While these in-
projects, and has carried out such projects across the highlands in at- centives may play a role in mediating uptake of methods by farmers,
tempts to transfer capacity to rural farmers and encourage the use of ADHENO does not directly facilitate market access or provide direct
these methods (Haregeweyn et al., 2015). Despite these broad efforts, monetary compensation for participation in trainings. We engage with
the policy has been sporadically successful; the likelihood of farmer the implications of these incentives in the discussion below. In a se-
adoption remains highly influenced by geography, socioeconomic and parate program, ADHENO facilitates the establishment of exclosures on
environmental conditions, and perceptions of the risks posed by soil communally managed lands and leads annual work days to plant
erosion and climate change (Anley et al., 2007; Kato et al., 2011; seedlings in exclosures.
Shiferaw and Holden, 1998). Thus, careful selection of geography for We sought to understand the impact of ADHENO’s farmer training
intervention, as well as more locally based needs assessments and program in addressing the economic and environmental challenges to
community engagement approaches may lead to greater rates of use of rural development in the two kebeles in which ADHENO has operated
such practices (Kato et al., 2011; Nigussie et al., 2017). the longest (Fig. 1). Specifically, we evaluated the degree to which
Farmer training programs and field schools are increasingly re- ADHENO’s approach is effective at engendering the use of conservation
cognized for their efficacy in achieving smallholder agricultural im- farming, addressing poverty, and mitigating land-use intensification. To
provements in areas such as soil and water conservation (Anley et al., do so, we utilized a mixed methods approach, combining an econo-
2007; Owenya et al., 2011), integrated pest management (Clausen metric household survey for propensity score matching with remote
et al., 2017; Jørs et al., 2016; Settle et al., 2014), and nutrient appli- sensing analysis of fine spatial resolution satellite imagery. We asked:
cation (Pan et al., 2017). These programs have been evaluated along (a) did farmers use practices taught in trainings; (b) did these strategies
many metrics, including adoption of farming techniques, changes in lead to improved agricultural productivity or livelihoods; and (c) how
livelihoods, and impacts on soil health and ecosystem services (Chindi did land-use intensity change in areas with more participation in farmer
et al., 2017; Davis et al., 2012; Feder et al., 2003; Takahashi and Todo, trainings?
2012). Yet, despite the intertwined nature of the environmental and
socioeconomic challenges facing farmers in the Ethiopian highlands, 2. Methods
assessments that interrogate the impacts of farmer trainings and field
schools in both dimensions are rare. If a program is successful in fos- 2.1. Description of the study area
tering implementation of conservation farming practices, will it also
improve social wellbeing and promote sustainable intensification on This study was undertaken in Goshebado and Metkoriya kebeles,
less agricultural land, enabling the rehabilitation of ecosystem services located in the North Shewa Zone of the Amhara Region of Ethiopia. We
across partially revegetated landscapes (Adimassu et al., 2016; Perfecto selected these kebeles because ADHENO has implemented farmer
and Vandermeer, 2008; Pretty and Bharucha, 2014)? To assess poten- trainings for over ten years in these areas, making them the longest-
tial for scalability and determine whether these programs can effec- standing project areas among ADHENO’s portfolio. The two kebeles are
tively advance development goals, it is important to build a holistic populated by 1159 and 540 households, respectively. The primary
understanding of their impacts. source of income for nearly all households is agriculture, livestock
Here, we evaluate the socioeconomic and land-use impacts of rearing, or a combination of the two. Most farmers grow a mix of cereal
farmer trainings focusing on soil and water conservation practices and pulse crops, supplemented by some vegetable cultivation.
conducted by a non-governmental organization (NGO). The NGO, Topographically, Goshebado and Metkoriya are characterized by a
ADHENO Integrated Rural Development Association (hereafter large central plateau flanked by steep escarpments leading down into
ADHENO), has operated in several rural kebeles in the North Shewa river valleys. Agriculture is mostly carried out on the central plateaus,
Zone, which lies in the central highlands of Ethiopia, since 2003 though some areas of the escarpment are also dedicated to farming.
(Fig. 1). Kebeles represent the smallest administrative unit in Ethiopia, Villages and homesteads are sparsely distributed across each kebele.
though it should be noted that large kebeles may also be split into sub- The elevation in the kebeles ranges between 2000 m to 2800 m. Total

2
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

Fig. 1. Map of Goshebado and Metkoriya kebeles, within the North Shewa Zone in Ethiopia.

annual rainfall is approximately 1147 mm per year (Harris et al., 2014), market town, distance to the farthest farm plots, elevation, slope, land
distributed between two rainy seasons; the short rains fall between holdings, household size, and prevalence of female heads of house-
February and April, and the long rains fall between June and Sep- holds. We conducted the same pairwise analysis to compare for parti-
tember. Farmers plow and sow their plots at the onset of the rainy cipating and non-participating households, since many farmers parti-
seasons, and harvest at different times throughout the dry season cipate in ADHENO trainings despite living in villages where no
months, between September and November, depending on the crop. trainings were hosted. In this comparison, we found that the only sig-
The soil is primarily comprised of clayey vertisols (Hengl et al., 2017). nificant difference between groups was that participating households
Goshebado and Metkoriya are connected to a nearby city, Debre have slightly more household members than non-participating house-
Berhan, by one unpaved road that runs across the central plateau holds.
connecting the two kebeles. These preliminary results suggest that ADHENO staff have not sig-
nificantly biased their selection of villages or households to participate
in trainings. Nonetheless, because of large within-group variation for
2.2. Evaluation challenges participants and non-participants along demographic and geographic
variables, we used a propensity score matching (PSM) approach for the
A key challenge to evaluating a farmer training program such as impact evaluation in order to assess differences between similar in-
ADHENO’s, in which the program was initially designed and im- dividual households in both groups. We also used robustness tests to
plemented without an evaluation plan in place, is the potential for se- assess the sensitivity of our findings to unobserved variables that may
lection bias. Where NGO staff make subjective decisions on villages for have biased ADHENO staff selection of villages and participants. This
hosting trainings and farmers to invite to trainings, there is a possibility approach is described in more detail below.
that these decisions are informed by socioeconomic and geographic
factors that can also predict the outcomes we hope to link with farmer
trainings in this evaluation. These factors include distance to market, 2.3. Household survey collection & analysis
gender and age of household head, level of education, household size,
land holdings, and – given the highly variable terrain – elevation and 2.3.1. Survey design
slope. Following the model of Haile et al. (2017), we addressed the To test the degree to which ADHENO activities have affected live-
possibility of selection bias by conducting pairwise comparisons to lihoods, we administered a quantitative and adaptive household survey
determine whether there are indeed differences along these variables to collect information on demographics, farming strategies, crop yields,
between program participants and non-participants. land holdings, income, education, and health. Surveys were collected
We first compared village-level means for the above variables be- by a team of five Ethiopian enumerators who were native speakers of
tween ADHENO-targeted villages and non-targeted villages, using our Amharic, the primary language of all Goshebado and Metkoriya re-
household survey described in detail below. After correcting p-values sidents. Surveys were collected using Qualtrics Survey Software on
for multiple comparisons using Bonferroni error correction (Haile et al., Android smartphones. The enumerators had no previous connection to
2017), we found there were no differences between targeted and non- ADHENO, and no incentive was offered to survey participants.
targeted villages in terms of average level of education, distance to the We used an adaptive survey tool that was structured to prompt

3
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

follow-up questions depending on the activities in which respondents the same variables as Panel 1, and added the following: village-level
had participated, and farming methods respondents used. Surveys means for slope, elevation, and distance to the market town; and sub-
consisted of between 25 and 127 questions and lasted between 15 and kebele of residence.
60 min, depending on the level of participation in ADHENO programs For both panels, we constructed a probit regression model to predict
and the variety and extent of the respondents’ farming activities. Survey the likelihood that each individual falls in the treatment group based on
respondents that had participated in ADHENO trainings were redirected the aforementioned variables, and matched individuals with similar
to answer additional questions about their perceptions and experiences scores in both groups for the PSM analyses. The primary output statistic
with those programs. The survey tool was initially designed in English from our PSM models is the average treatment effect on the treated
and was translated into Amharic by an English professor at Debre (ATT score), which represents the average effect on those who received
Berhan University. The translation was then verified by another native the treatment (Li, 2012b). We also conducted Rosenbaum’s bounds tests
Amharic speaker who is fluent in English. Prior to implementation, we for sensitivity to hidden bias for each PSM model to account for any
field tested the survey with farmers and made minor revisions to the variables omitted from the matching analysis that may have mediated
survey tool to improve ease of use for enumerators and address com- participation in farmer trainings.
munity-specific issues with the survey questions. The survey tool was We first asked: do farmers that participate in trainings use the
granted exempt status by the Institutional Review Board at the practices taught? We created a dummy variable for use of each farming
University of Michigan. method taught by ADHENO (digging trenches, terracing, planting ve-
getation to stabilize soil, use of improved seeds, intercropping, fertilizer
2.3.2. Survey study population and sampling method application, and irrigation), and tested the difference in rate of use
The survey was administered between June and August 2017. between treatment and control groups. We also constructed non-PSM
Overall, we collected surveys from 449 of the 1700 households in logistic regressions to test whether likelihood of using each conserva-
Metkoriya and Goshebado. At that time, all survey respondents were tion farming method is influenced by attendance at multiple trainings,
living in the Goshebado or Metkoriya kebeles. To ensure adequate re- controlling for the variables used in the PSM analysis. We then asked: is
presentation of households, we grouped all villages in both kebeles into participation in farmer trainings associated with improvements in crop
those that had and had not been targeted for farmer trainings. Then, we productivity? For this, we calculated yield per hectare for each crop
randomly selected villages from each category for random sampling of using self-reported metrics for crop productivity and area dedicated to
households within these villages. Households recruited in this way ac- each crop, and compared yield differences between the treatment and
counted for 265 out of 449 survey respondents. The majority of re- control groups. Lastly, to determine whether participation in farmer
spondents were surveyed in the central market town of the kebele, at trainings is associated with improved agricultural livelihoods, we
community events such as religious festivities, or in the central village compared income from farming between treatment and control groups,
of sub-kebeles. Respondents were recruited in one of the following using self-reported income from farming. Stata 15.0 was used for PSM
ways: a) sub-kebele leaders gathered heads of households to meet at a analysis and RStudio 1.1.442 was used for non-PSM logistic regressions.
central location in their village or the kebele market town, b) enu-
merators visited households in villages selected for random sampling, 2.4. Land-use intensity
and c) respondents were recruited by the enumerators while walking on
the only road through the kebeles. We also included a convenience To determine whether village-level participation in ADHENO
sample to recruit an additional 184 respondents from the two kebeles to farmer trainings was associated with land-use intensification over time,
participate in our survey. These additional respondents were recruited we used fine spatial resolution (< 2 m pixels) satellite images to classify
using the same methods as above, but were from villages that were not land-use/land-cover (LULC) over time and modeled changes in land-use
selected in the random sample. No specific households were targeted or intensity within village areas. This analysis was limited to Goshebado
rejected for participation in the survey. for multiple reasons. First, we collected more surveys in Goshebado, so
we were better able to test associations between farmer training par-
2.3.3. Variable management and analysis ticipation and land-cover change in that area. Second, we relied on
Because we were primarily interested in understanding how parti- ADHENO field staff and community members to identify village loca-
cipation in farmer trainings affected implementation of soil and water tions on Google Earth, and these informants were only able to con-
conservation techniques and household income, we conducted a pro- fidently identify Goshebado villages.
pensity score matching (PSM) analysis of household surveys. PSM is a Due to extreme patchiness of the landscape, characterized by
quasi-experimental statistical method used to estimate the effect of a smallholder farm plots interspersed with eucalyptus woodlots and
treatment by matching individuals in the treatment group with similar patches of pasture and shrubland on escarpments, we used an object-
individuals in the control group, approximating a counterfactual con- based approach for this analysis. Multispectral images were available
trol group (Li, 2012a; Rosenbaum and Rubin, 1983). Here, we con- from December 2007 and November 2016, providing a nine-year
structed a binary variable for participation in any number of trainings window over which to assess change in land-use intensity. November
and participation in no trainings, and considered the former group as and December fall during a dry season harvest period, creating sig-
the “treatment” and the latter as the “control.” nificant heterogeneity in the appearance of agriculture and vegetation
Due to varying availability of geographic information between the across the landscape. Both images provided 100 percent coverage of
two kebeles, we conducted the PSM analysis in two panels. We were Goshebado, a 72 km2 area, and had 0 percent cloud cover. All image
able to record village locations for calculating geographic variables for analysis was carried out in ERDAS Imagine and ArcGIS Desktop 10.5.
households in Goshebado, but were unable to do so for Metkoriya.
Additionally, we recorded sub-kebeles for survey respondents from 2.4.1. Image preprocessing
Goshebado, but were unable to ascertain sub-kebeles for respondents The 2007 image was captured by the Korea Multi-Purpose Satellite
from Metkoriya. Thus, we constructed the Panel 1 PSM analysis with (KOMPSAT-2) system, which contains blue, green, red, and near-in-
survey data from both kebeles (n = 449), while Panel 2 represented a frared bands at 4 m resolution, as well as a 1 m panchromatic band. The
more robust PSM analysis with surveys from Goshebado only 2016 image was captured by WorldView-2, and comprised of four
(n = 339). In Panel 1, we used the following variables to match treat- multispectral bands at 1.3 m resolution and a 0.5 m panchromatic band.
ment and control groups: land holdings, age and gender of the head of We orthorectified both images using a 30 m SRTM digital elevation
household, household size, distance to the farthest plot, the terrain of model (DEM) and pan-sharpened the 2007 multispectral image to a
plots, level of education, and kebele of residence. In Panel 2, we used resolution of 1 m, then resampled the result to 1.3 m pixels to

4
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

Table 1
Land-use intensity and associated land-cover classifications and descriptions, adapted from land-cover/land-use classification systems in Delelegn et al. (2017) and
Eggen et al. (2016).
Land-use intensity level Land-cover class name Description

1. Low intensity Woodland Land not used for agriculture, with vegetation cover between 20-100%
Water All sources of open water – primarily found in rivers in the valleys delineating Goshebado kebele from neighboring kebeles
2. Low-medium intensity Eucalyptus Plantation Identified by orthogonal tracts of dense vegetation, surrounded by very little vegetation or settlements
3. Medium intensity Shrubland/ grassland Land not used for agriculture, with vegetation cover between 1-20%
Bare soil Land not used for agriculture with 0% vegetation cover
4. High-medium intensity Smallholder Agriculture Patchwork of farm plots with low density of settlements
5. High intensity Rural settlement High density of settlement and impervious surface
Other No data Shadows and other areas for which land-cover was impossible to discern

standardize pixel size between the two images. We conducted accuracy assessments for the classified images
(Fig. 3) using a unique set of test pixels for each image that were
manually classified by a trained observer, using the original satellite
2.4.2. Image classification image for reference. Test pixels were selected using a random stratified
We classified LULC for each image using an eight-category classi- sample, with each class receiving a minimum of 30 test pixels. We
fication system based on a scheme developed for the Ethiopian high- calculated producer and user accuracy for each LULC class; overall
lands by Eggen et al. (2016) and refined for the Goshebado context with kappa was 0.77 for the 2007 image, and 0.86 for the 2016 image.
on-the-ground observations (Table 1). Specifically, we added a class for
Eucalyptus woodlot due to the livelihood importance of woodlots in
this region and ecological differences between exotic eucalyptus plan- 2.4.3. Measuring change in land-use intensity
tations and indigenous forest, and removed the wetland class due to the In order to measure land-use intensification over time, we first
lack of wetlands. Images were first processed to identify spectral sig- characterized each of the LULC classes in terms of their land-use in-
natures of LULC classes at the pixel level, segmented to objects which tensity based on previous studies of historic land-cover and land-use
were characterized by shape, size, pixel reflectance, and texture, then impacts on soil stability and nutrient levels (Table 1). Geological evi-
classified using ERDAS Objective, an object-based image analysis dence reveals that the Ethiopian highlands were forested before an
package in ERDAS Imagine. We classified segments by applying an increase in agricultural and pastoral pressure upon the landscape be-
adapted machine-learning approach previously used to extract trees ginning 2000–3000 years ago (Darbyshire et al., 2003), suggesting that
from WorldView-2 satellite imagery (Chepkochei, 2011), described in woodland reflects the lowest intensity of land-use among the LULC
more detail below. Prior to the extraction of classes in ERDAS Objec- classes present in Goshebado. We classified Eucalyptus woodlots as low-
tive, we digitized the rural settlement and masked it out of the image, as medium land-use intensity. Woodlots have higher levels of soil organic
there is only one densely settled area in Goshebado. carbon and soil stability relative to agricultural and pastoral land
We extracted LULC classes using an iterative model-building pro- (Delelegn et al., 2017); however, woodlots have more acidic soil com-
cess. For each target LULC class, we identified a set of training pixels pared to forests, and periodic harvests of trees contribute to lower
that fell within the class and another set of pixels that fell outside the nutrient levels compared to forests (Liang et al., 2016). Shrubland/
target class as background pixels for comparison. We trained a Single grassland and bare soil were designated as medium land-use intensity,
Feature Probability (SFP) pixel classifier, which calculated the prob- as they likely reflect areas frequently used for livestock grazing. Grazing
ability that all pixels in the images were in the target LULC class. We represents frequent low-intensity nutrient removal from soils, resulting
then performed an unsupervised object-based segmentation using the in intermediate soil stability and nutrient levels that are between
Full Lambda Schedule Image Segmentation to specify desired levels of woodland and farmland (Delelegn et al., 2017; Yimer et al., 2007).
homogeneity of size, texture, shape, and pixel reflectance for output Agricultural lands were designated high-medium intensity, due to the
segmentation polygons. We used a 30 m minimum mapping unit for pressure on soil organic matter and soil stability caused by the farming
segmentation. Following segmentation, the SFP classifier was used to of annual crops (Delelegn et al., 2017). Rural settlement represents the
calculate the probability that each segment fell within the target class most intense land-use, given the propensity for urbanization to frag-
based on spectral information. A second set of probabilities was cal- ment ecosystems and add to impervious surface (Lambin et al., 2001).
culated using vector geometry parameters for all polygons, including Changes in land-use intensity levels were determined by comparing
area, perimeter, rectangularity, and eccentricity of polygon outlines, the land-use intensity of LULC classes as observed in 2007 with land-use
and trained based on a set of training objects that fell within the target intensity for 2016. Changes were summarized by calculating two binary
class. The probability from pixel and vector-based calculations were variables for each pixel: increase in intensity (pixels having no change
then combined, providing a total probability score between 0 and 1 that and decreasing intensity = 0; pixels having increasing intensity = 1)
each object fell within the target LULC class. We then applied a filter to and decrease in intensity (no change and increasing intensity = 0; de-
remove objects with probability scores falling below a threshold which creasing intensity = 1).
varied from 0.3 – 0.6 depending on the model. Prior to building a model
to extract the next class, we masked classified objects out of the image 2.4.4. Variable management and analysis
to ensure no features were classified twice. We repeated this process to Due to a lack of spatially explicit data for farm plots managed by
sequentially extract all LULC classes for both images, until there were each surveyed household, we modeled change in land-use intensity at
few remaining unclassified pixel clusters. We then manually classified the landscape level, with individual villages as the unit of analysis. We
all remaining unclassified polygons larger than the minimum mapping aggregated land-use intensity data and household surveys to the village
unit and merged all LULC classes into a single file that was processed to level to create the variables for this analysis, and calculated several
remove topological errors. We used this process to first classify water, spatial covariates. In consultation with ADHENO staff, we identified the
followed by shadow/no data, eucalyptus woodlot, agriculture, wood- locations of villages in Goshebado (n = 37) using Google Earth, and
land, shrubland, and bare soil, and determined this order by prioritizing generated Thiessen polygons using village coordinates to approximate
the LULC classes that the algorithm was more adept at identifying village boundaries.
(Fig. 1). For each village, we calculated the rate of increase and decrease in

5
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

land-use intensity. From surveys, we calculated village-level rate of training sessions predicts the implementation of soil conservation
participation in ADHENO farmer trainings and seedling plantation farming strategies. The difference-in-means value represents the rate by
programs, and average land holdings, income, and level of education. which the treatment group is more likely to use a particular method
We used census sheets to determine the number of households in each than the control group; for example, farmers who participate in any
village. We also used a 30 m SRTM DEM to calculate spatial covariate ADHENO training are 31–44 percent more likely to practice inter-
rasters for slope as percent rise, path distance from the central market cropping than non-participants.
town, and Euclidean distance from ADHENO exclosure sites. We used The Rosenbaum bounds test, used to determine the sensitivity of the
these rasters to calculate the average for each of these covariates within results to hidden bias or unconsidered factors mediating participation
village Thiessen polygons. The result of this process was a dataset as- in the treatment group, are represented by the Gamma value (Γ). These
sociating village-level changes in land-use intensity with demographic values range from 2.45 for terracing in Panel 2 to 21.85 for use of
and spatial variables. trenches in Panel 2. This can be interpreted to mean that effects of
We used linear models to test whether there was a significant linear training would have to differ from the predicted effect (as a result of
relationship between the rate of participation in farmer trainings and unobserved variables) by 2.45 to 21.85 times for unobserved variables
the rate of increase and decrease in land-use intensity at the village to invalidate the significance of the relationship between farmer
level. To avoid undersampling bias, we filtered out villages for which trainings and use of terracing or trenches, respectively. The effect is
we had surveyed fewer than 15% of households, leaving 28 villages for considered significant if Γ = 1 and demonstrates increasing robustness
the analysis. We controlled for number of households, average slope, of the PSM test as Γ increases (Becker and Caliendo, 2007).
distance to the town, education, income, land holdings, and participa- We also conducted non-PSM logistic regressions, controlling for the
tion in the ADHENO exclosure and seedling plantation program. Linear demographic and geographic variables used for matching above, to test
models were first tested with variance inflation factors to check for whether ADHENO-trainees were more likely to implement farming
collinearity among covariates, and Shapiro-Wilks tests were performed methods if they attended more trainings. We found that across both
on model residuals to test the assumption of homoscedasticity. No panels, the odds of implementing trenches and intercropping was sig-
covariates were found to be collinear, and assumed homoscedasticity nificantly higher for farmers that attended multiple trainings, while the
held. We used RStudio Version 1.1.442 to conduct the statistical ana- number of trainings attended did not influence the rate of planting
lysis. vegetation to stabilize soil (Table 4). At a 90% confidence interval,
farmers that attended more trainings were more likely to use trenching.
3. Results
3.1.3. Crop productivity
3.1. Farmer survey For all crops, there was no significant difference in crop productivity
between treatment and control groups in both PSM panels (Table 5).
3.1.1. Descriptive results The sample size of farmers growing oil beans was not sufficiently large
Of the 449 households surveyed, 100% cited farming as their pri- to perform a matching analysis.
mary or secondary source of income. All but one household reported
their ethnicity as Amharic. The majority of survey respondents (77%) 3.1.4. Income from farming
were male. Seventy-five percent of households reported participating in Regressing the dummy variable representing attendance at any
one or more ADHENO farmer training (Table 2). farmer training with self-reported income from farming over the pre-
vious year showed that attendance at farmer trainings is associated with
3.1.2. Use of soil and water conservation farming practices higher income from farming in both PSM panels (Table 6). Results from
Analysis of the relationships between participation in ADHENO this propensity score matching analysis show that the treatment group
farmer trainings and self-reported use of the farming practices taught in (i.e., those attending a training session) had on average an income that
the trainings yielded a significant effect for four out of seven methods: was higher by 4995 Ethiopian Birr (ETB), or about 217 US Dollar (USD)
terracing, intercropping, use of vegetation to stabilize soil, and tren- compared to the control group for the Panel 1 PSM analysis across
ches. These results were consistent across both PSM panel analyses Goshebado and Metkoriya kebeles. In Goshebado alone, represented by
(Table 3). All ATT t-stat coefficients were positive, with the exception of Panel 2, the difference between treatment and control group farming
irrigation for both panels and fertilizer for Panel 2. The results were not incomes was 6265 ETB, or approximately 272 USD.
significant for fertilizer, use of improved seeds, or irrigation. The po-
sitive significant relationship between participation in any training and 3.2. Land-use intensity
the use of terracing, intercropping, use of vegetation to stabilize soil,
and trenches indicates that attendance in any of the ADHENO farmer 3.2.1. Descriptive results
On average, 63.38% of the land-use within delineated village areas
Table 2 remained the same over time (n = 28, sd = 6.72; Figs. 2 and 3). In
Descriptive results for key variables of interest, and demographic and geo- contrast, across villages, land-use intensity increased in an average of
graphic variables used for matching analysis. 14.47% of village area (n = 28, sd = 6.98; Figs. 2 and 3), and decreased
Variable Mean Standard Median in an average of 21.85% of village area (n = 28, sd = 7.94; Figs. 2, and
Deviation 3). Meanwhile, participation in farmer training sessions was high, with
an average of 70% participation in at least one session across villages,
Farmer Trainings Attended in Previous 5 4.47 3.57 5
though there was substantial variation between villages (n = 28,
Years
Income from Farminga 8913.04 7745.66 8000 sd = 18.56; Figs. 2 and 3).
Land holdings in Timadsb 6.24 2.93 6
Male Head of Household Age 51 13.08 50 3.2.2. Predictors of land-use change at the village-level
Female Head of Household Age 44 11.98 44 Linear models of rates of increase and decrease in land-use intensity
Household Size 5.14 1.99 5
indicated that participation in farmer trainings was not significantly
a
Farmers reported their income in Ethiopian Birr; the exchange rate to US associated with rate of land-use change at the village level (increasing
Dollar was 23:1 at the time of this study. intensity: n = 28, p = 0.89; decreasing intensity: n = 28, p = 0.14).
b
Farmers reported land-holdings in the local unit of area measurement, the However, land-use change was significantly associated with other
timad. One timad is equivalent to 0.25 ha (Shiferaw et al., 2001). variables in the model (increasing intensity: R2 = 0.68; decreasing

6
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

Table 3
Results of PSM probit regression of Farmer Trainings Attended (0/1) & Use of a Method (0/1) and Rosenbaum bounds test (listed as Γ). Panel 1 included individuals
in both Metkoriya and Goshebado, and used only demographic variables for matching. Panel 2 included survey respondents from Goshebado, and used both
geographic and demographic variables to match. The ATT t-statistic denotes statistical significance of the effect, with a value of 1.95 or greater indicating 95%
confidence. Variables in bold typeface indicate a significant relationship between attending farmer trainings and use of the method.
Farming method

Fertilizer Terracing Improved Seeds Soil-stabilizing Vegetation Inter-cropping Trenches Irrigation

Panel 1
Treatment mean 0.976 0.895 0.559 0.594 0.521 0.283 0.059
Control mean 0.962 0.521 0.404 0.136 0.077 0 0.209
SE 0.064 0.145 0.132 0.115 0.082 0.027 0.138
ATT t-stat 0.22 2.58*** 1.19 3.95*** 5.38*** 10.61*** −1.09
Gamma <1 3.8 1.15 3.75 3.35 6.8 <1

Panel 2
Treatment mean 0.97 0.885 0.555 0.59 0.465 0.295 0.065
Control mean 1 0.52 0.39 0.115 0.155 0 0.08
SE 0.012 0.163 0.165 0.131 0.091 0.032 0.137
ATT t-stat −2.48 2.24*** 1 3.63*** 3.40*** 9.13*** −0.11
Gamma <1 2.45 1.3 5.75 3.9 21.85 <1


Note: 90% CI, * 95% CI, ** 99% CI, *** 99.9% CI.

Table 4 method. However, there were no differences in crop productivity be-


Non-PSM logistic regression demonstrating the influence of number of trainings tween attendees and non-attendees, and higher rate of participation in
attended on the odds of using a farming method, controlling for demographic farmer trainings was not associated with changes in land-use intensity
variables. This test was limited to farming methods found to be used sig- at the village level. We discuss the implications of these findings below.
nificantly more frequently by ADHENO training participants, and included only For four of seven training topics – terraces, planting vegetation to
farmers that had participated in at least one training. Panel 1 included survey
stabilize soil, intercropping, and digging trenches – farmers that at-
respondents from Goshebado and Metkoriya, while Panel 2 included survey
tended trainings utilized those practices at significantly higher rates
respondents only from Goshebado and controlled for geographic variables.
Variables in bold typeface indicate a significant association between number of than farmers that did not attend trainings, after matching farmers with
trainings attended and use of the farming method. similar demographic and geographic characteristics from both groups.
ADHENO-trained farmers were approximately 30% to 48% more likely
Farming method
to use these methods than non-participating farmers. These results were
Terracing Soil-stabilizing Inter- Trenches consistent across the two PSM panels, highly robust to omitted variable
Vegetation cropping bias, and were achieved even though our treatment variable didn’t
discriminate between types of training. Furthermore, for each addi-
Panel 1
tional training attended, the odds that an ADHENO training participant
Odds ratio 1.368*** 1.017 1.111*** 1.095‡
Observations 335 335 335 335 built terraces increased by 37%, and the odds of using intercropping
df 313, 322 313, 322 226, 238 226, 238 increased by 11%.
AIC 236.38 427.60 328.35 287.88 Nearly all farmers used fertilizer, very few used irrigation, and there
Panel 2 was a non-significant 15% difference between groups in rate of use of
Odds ratio 1.283** 1.049 1.111* 1.094‡ improved seeds. For fertilizer and irrigation, our findings reflect the fact
Observations 251 251 251 251 that agricultural extension agents provide access to inorganic fertilizers
df 226, 238 226, 238 226, 238 226, 238
at the weekly market, and that irrigation requires specific topographical
AIC 211.40 330.15 328.35 287.88
farm conditions and investment into expensive equipment. Our finding
‡ that participants and non-participants used improved seeds at similar
Note: 90% CI, * 95% CI, ** 99% CI, *** 99.9% CI.
rates is unexpected, since ADHENO provides access to improved seeds
intensity: R2 = 0.55; Table 7). Increasing land-use intensity was posi- and farm tools at farmer trainings, yet participating farmers are more
tively associated with slope, such that each percent increase in slope likely to use methods that require landscape modification, such as
predicted a 0.46 percent increase in the rate of land-use intensity (p = terracing and trenches. Limited access to resources can constrain
0.001). Decreasing land-use intensity was positively associated with adoption of certain costly farming practices (Adimassu et al., 2016),
average land holdings and distance from ADHENO-facilitated ex- thus one might expect provisioning improved seeds and farm tools to
closures, such that each additional hectare predicted an additional 2.74 facilitate uptake of all related methods. We may draw two potential
percent decrease in land-use intensity (p = 0.049), and each kilometer insights from this result: a) agricultural extension agents carry out se-
of distance from exclosures predicted an additional 2 percent of de- parate programs to distribute improved seeds, but not farm tools; or b)
crease in land-use intensity (p = 0.03). Shapiro-Wilks tests indicated planting improved seeds requires lower investments of time, labor, and
that the assumption of homoscedasticity held (increasing intensity: p = land compared to landscape modifications such as trenches and ter-
0.52; decreasing intensity: p = 0.08). races, making it an easy technology to adopt.
Interestingly, while the methods used at similar rates between
treatment and control groups can improve crop yield and resilience, the
4. Discussion
practices used more frequently by ADHENO-trained farmers are more
oriented to soil conservation (Haregeweyn et al., 2015). Constructing
This study of a soil conservation farmer training program in the
terraces, digging trenches, and planting vegetation to stabilize soil are
Ethiopian highlands reveals that farmers that attended trainings used
all costly practices, as they may require land and labor that could
several of the methods taught at higher rates and had higher income
otherwise be put towards farm productivity (Adimassu et al., 2016).
from farming compared to non-attendees. For two farming methods,
Additionally, though intercropping represents potential yield
participation in more trainings predicted greater likelihood of using the

7
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

Table 5
Results of PSM probit regression of Farmer Trainings Attended (0/1) & Crop Productivity in kilograms per hectare and Rosenbaum bounds test (listed as Γ). Panel 1
included individuals in both Metkoriya and Goshebado, and used only demographic variables for matching. Panel 2 included survey respondents from Goshebado,
and used both geographic and demographic variables to match. The ATT t-statistic denotes statistical significance of the effect, with a value of 1.95 or greater
indicating 95% confidence. Variables in bold typeface indicate a significant association between training attendance and crop yield.
Crop

Wheat Beans Teff Barley Chickpeas Lentils

Panel 1
Treatment mean yield (kg/ha) 990.870 867.552 714.890 845.147 611.692 559.751
Control mean yield (kg/ha) 948.416 974.778 651.216 582.161 664.179 296.285
Difference in means 42.455 −107.226 63.674 262.985 −52.488 263.466
SE 228.539 433.701 163.956 286.351 190.674 307.360
ATT t-stat 0.19 −0.25 0.39 0.92 −0.28 0.86
Γ <1 <1 <1 <1 <1 <1

Panel 2
Treatment mean yield (kg/ha) 944.905 847.517 639.931 836.173 571.147 538.803
Control mean yield (kg/ha) 1153.571 1067.563 740.215 659.289 527.957 443.373
Difference in means −208.666 −220.047 −100.284 176.884 43.190 95.430
SE 228.833 462.465 184.269 522.553 355.501 390.725
ATT t-stat −0.91 −0.48 −0.54 0.34 0.12 0.24
Γ <1 <1 <1 <1 <1 <1


Note: 90% CI, * 95% CI, ** 99% CI, *** 99.9% CI.

Table 6 begin to increase (Badgley et al., 2007; Pittelkow et al., 2015). Thus,
Results of PSM probit regression of Farmer Trainings Attended (0/1) & Income while farmers that attended trainings had similar crop yields to other
from Farming (Ethiopian Birr) and Rosenbaum bounds test (listed as Γ). Panel 1 farmers at the time we collected household surveys, productivity be-
included individuals in both Metkoriya and Goshebado, and used only demo- tween the two groups may diverge over time.
graphic variables for matching. Panel 2 included survey respondents from While attendance at farmer trainings predicts use of conservation
Goshebado, and used both geographic and demographic variables to match.
practices on farmers’ plots, we did not find an association between the
The ATT t-statistic denotes statistical significance of the effect, with a value
program and landscape-level patterns in land-use intensity change.
≥1.95 indicating 95% confidence. Bold typeface indicates a significant differ-
ence in income (ETB) between treatment and control groups.
Even so, land-use intensity appears to be in flux, with some areas in-
tensifying and other areas decreasing in intensity over time. Our in-
Income from farming conclusive findings may reflect a lack of statistical power; without
Panel 1 Panel 2
spatial coordinates for farm plots managed by each surveyed house-
Treatment mean 10228.821 9998.779 hold, we were unable to test the relationship between training parti-
Control mean 5233.333 3733.14 cipation and land-use change at the household level. The murky re-
Difference in means 4995.488 6265.639 lationship between land-use intensity and farmer trainings is also likely
SE 1412.948 1979.619
related to land tenure policy in Ethiopia. The government owns all land
ATT t-stat 3.54*** 3.17***
Γ 1.93 2.65 and leases land to farmers to use for short periods of time (Ege, 2017).
Unfarmed land in this area may be open grazing land, or may be co-

Note: 90% CI, * 95% CI, ** 99% CI, *** 99.9% CI. managed rangelands held by up to 20 households (I. Fikre, personal
communication, June 2017). The ADHENO exclosure program, not
reductions compared to planting the crops individually, it can also ef- studied here, brings together households that co-manage parcels of
ficiently restore soil nutrients and organic carbon while continuing to particularly erosion-prone rangeland to establish exclosures for re-
produce cereal cash crops (Agegnehu et al., 2006; Beedy et al., 2010). habilitating vegetation. Vegetation plays an important role in providing
Our finding that ADHENO-trained farmers use these methods more ecosystem services that support farmer livelihoods and native biodi-
frequently than non-participants strongly indicates that the trainings versity, through the maintenance of soil nutrients and water catchment,
enable farmers to overcome knowledge and resource constraints on sequestering carbon and conserving native plant species, highlighting
implementing soil conservation measures, supporting previous research the importance of improving vegetation structure in this area (CEFP,
on this topic (Adimassu et al., 2016). 2012; Tolessa et al., 2017; Yimer et al., 2007). However, there are few
Nonetheless, the higher use of conservation farming methods among ADHENO exclosures and they are small, leaving much of the land in
trained farmers is somewhat surprising, given the finding that crop valleys and on escarpments as open grazing lands. Insecure or unclear
yield is similar between trained and untrained farmers. Smallholder land tenure makes individual investments into reforestation and sus-
farmers are sensitive to the potential risks and gains of adopting new tainable land management practices difficult to achieve without larger
farming practices (Adimassu et al., 2016; Shiferaw et al., 2009), and are national policies supporting such activities (Legesse et al., 2018). As an
highly motivated to continue using sustainable land management NGO, ADHENO is not able to enact any policies to incentivize refor-
practices if the techniques are linked to yield improvements over time estation or reform land tenure in the North Shewa Zone, but may be
(de Graaff et al., 2008). However, even without measurable pro- able to impact land-use by offering trainings on livestock management
ductivity gains, the perceived benefits from implementing these prac- or mixed crop-livestock farming systems in the future.
tices can mediate their initial use and long-term adoption (Amsalu and Despite no difference in crop yield between farmers that attended
de Graaff, 2007; de Graaff et al., 2008), and it is possible that ADHENO and did not attend trainings, the former group reported far greater
trainings are highly effective at engendering this perception among annual income from agriculture compared to the latter. It should be
participants. Implementation of these practices may also be mediated noted that while crop yields found in this study are below the national
by farmer perceptions of the long-term benefits of soil conservation. average for smallholder farmers in Ethiopia (Taffesse et al., 2011), these
Alternatively, it may take years after the implementation of conserva- yields are within range of yields reported in previous studies of crop
tion practices for soil health to improve to a point where crop yields productivity in the Ethiopian highlands (Agegnehu et al., 2006; Amare

8
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

Fig. 2. Classified land-use/land-cover for 2007 and 2016, participation rate in farmer trainings, and change in land-use intensity over time in Goshebado kebele.
Land-cover classification was based on KOMPSAT-2 imagery from November 2007 and WorldView-2 imagery from December 2016. Village boundaries were
estimated as Thiessen polygons calculated from village centerpoints. Villages were excluded from analysis if fewer than 15% of households were surveyed.

et al., 2013; Getachew et al., 2008; Oicha et al., 2010; Vancampenhout However, farmer trainings on their own appear to be insufficient to
et al., 2006). Additionally, we matched participants with non-partici- mitigate soil degradation at the landscape level, and NGOs and other
pants based on overall land holdings and household size, and normal- practitioners may consider developing programs to leverage communal
ized crop yield by the area that each farmer dedicated to the particular land management and livestock-raising practices to rehabilitate de-
crop, making it unlikely that issues of land holdings would significantly graded lands.
affect this finding. This study also produces a useful roadmap for evaluating the socio-
There are several possible explanations this phenomenon. One environmental outcomes from development interventions in circum-
possibility is that ADHENO advises farmers on market access, or the stances where the program is implemented with no evaluation frame-
best times to sell crops in order to take advantages of inflection points work in place. Many small NGOs and social enterprises have innovative
in the crop market values. We also found that as a group, training and localized approaches to improving the livelihoods and resilience of
participants devoted more land to cultivating wheat and beans – the smallholder farmers. Yet these programs often lack the resources and
two most productive crops by weight in this study – than non-partici- capacity to establish evaluation frameworks and collect baseline data
pants. Additionally, Getachew et al. (2008) showed that farmers can while also implementing programs. It is critically important to under-
obtain a higher land-equivalent ratio for bean-wheat intercropping in stand the impacts of such programs in order to identify and potentially
the Ethiopian highlands, compared to cultivating either crop as a scale up highly effective models. By pairing a matching econometric
monoculture. If ADHENO advises farmers to plant these two crops in analysis with long-term satellite monitoring, we present a framework
combination, their yields would appear to be within range of other for future evaluators to assess the degree to which localized interven-
farmers, but they would gain far more income per unit of land. While tions can address the complex social and environmental challenges to
our survey asked whether farmers practiced intercropping, we did not establishing resilient livelihoods in the Ethiopian highlands.
collect data on which specific crops were planted together, and in what
volumes, so it is not possible for us to test this hypothesis. Nonetheless,
Declaration of competing interest
the large difference in income between similar farmers in treatment and
control groups is a strong indicator that ADHENO trainings play a role
None.
in generating agricultural income for participants, aligning with pre-
vious literature showing that farmer trainings and field schools can
effectively alleviate poverty among smallholder farmers. Competing interests

The authors declare they have no competing interests. Funding


5. Conclusions
sources were not involved in the decision to publish this research and
had no influence on the findings of the research.
This study illustrates that farmer training programs can effectively
promote the use of soil and water conservation techniques among
smallholder farmers in the Ethiopian highlands, while also improving Consent for publication
farmers’ abilities to generate income from crops. Given the conditions
facing farmers in the Ethiopian highlands and other erosion-prone All project participants provided their informed consent. All re-
areas, this form of intervention may have resounding impacts on search was performed in accordance with the policies of the University
household resilience to climate change, drought, and other shocks. of Michigan Internal Review Board.

9
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

Fig. 3. Distribution among villages of (a) participation in farmer trainings, (b) total area, (c) percent area with no land-use change, (d) percent area with land-use
intensification, and (e) percent area with land-use extensification.

Table 7 Funding
Linear model results of land-use change at the village level. Beta coefficients are
reported with significance levels, and standard errors in parentheses. This work was supported by a grant from the NASA Land Cover and
Dependent Variable Land Use (LCLUC) program (NNX15AD40G). Commercial satellite
imagery was provided through NASA’s NGA Commercial Archive Data
Rate of increase in Rate of decrease in (cad4nasa.gsfc.nasa.gov) under the National Geospatial-Intelligence
land-use intensity land-use intensity Agency’s NextView license agreement.
Participation in farmer −0.009 (0.063) −0.131 (0.087)
trainings
Participation in seedling −0.029 (0.168) 0.385 (0.230) Acknowledgments
plantation program
Average land holdings −1.488 (0.950) 2.73* (1.305)
The authors would like to express immense gratitude to the farmers
Average income 0.000 (0.000) 0.000 (0.000)
Average education 4.218 (3.376) −0.654 (4.636) of Goshebado and Metkoriya for sharing their time and insights for the
Number of households 0.067 (0.072) −0.095 (0.098) purposes of this evaluation. This work was made possible by the
Slope 0.463** (0.120) −0.073 (0.164) University of Michigan, the William Davidson Institute, and ADHENO
Distance from town −0.007 (0.004) 0.004 (0.006)
Integrated Rural Development Association. The authors would like to
Distance from ADHENO −0.029 (0.168) 0.003* (0.001)
exclosures
thank the ADHENO staff, particularly Rahel Hiletework, Isayias Fikre,
Constant 10.912 (9.521) 0.657 (13.074) and Ayele Beyene for their support during fieldwork. The authors are
Observations 28 28 also grateful for the translation, interpretation, and survey enumeration
R2 0.680 0.554 services of Mezemir Girma, Hana Mekonen, Tesfa Hailu, Mahlet Aklil,
Adjusted R2 0.520 0.332
Nahom Desalegn, and Dawit Girma. Finally, the authors are apprecia-
Residual Std. Error (df = 18) 4.762 6.538
F Statistic (df = 9; 18) 4.243 2.488 tive of the technical support offered by the University of Michigan
Office for Consulting for Statistics, Computing & Analytics Research and

Note: 90% CI, * p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.001. Shannon Brines of the University of Michigan Environmental Spatial
Analysis Laboratory for critical feedback on analyses. The views ex-
pressed here are the authors’ alone.

10
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

References Jørs, E., Konradsen, F., Huici, O., Morant, R.C., Volk, J., Lander, F., 2016. Impact of
training Bolivian farmers on integrated pest management and diffusion of knowledge
to neighboring farmers. J. Agromedicine 21, 200–208. https://doi.org/10.1080/
ADHENO, 2015. ADHENO’s Second Strategic Plan (2016-2020). Addis Ababa, Ethiopia. 1059924X.2016.1143428.
Adimassu, Z., Langan, S., Johnston, R., 2016. Understanding determinants of farmers’ Kato, E., Ringler, C., Yesuf, M., Bryan, E., 2011. Soil and water conservation technologies:
investments in sustainable land management practices in Ethiopia: review and a buffer against production risk in the face of climate change? Insights from the Nile
synthesis. Environ. Dev. Sustain. https://doi.org/10.1007/s10668-015-9683-5. basin in Ethiopia. Agric. Econ. 42, 593–604. https://doi.org/10.1111/j.1574-0862.
Agegnehu, G., Ghizaw, A., Sinebo, W., 2006. Yield performance and land-use efficiency of 2011.00539.x.
barley and faba bean mixed cropping in Ethiopian highlands. Eur. J. Agron. 25, Lambin, E.F., Turner, B.L., Geist, H.J., Agbola, S.B., Angelsen, A., Folke, C., Bruce, J.W.,
202–207. https://doi.org/10.1016/j.eja.2006.05.002. Coomes, O.T., Dirzo, R., George, P.S., Homewood, K., Imbernon, J., Leemans, R., Li,
Amare, T., Terefe, A., Selassie, Y.G., Yitaferu, B., Wolfgramm, B., Hurni, H., 2013. Soil X., Moran, E.F., Mortimore, M., Ramakrishnan, P.S., Richards, J.F., Steffen, W., Stone,
properties and crop yields along the terraces and toposequece of Anjeni Watershed, G.D., Svedin, U., Veldkamp, T.A., 2001. The Causes of Land-Use and Land-Cover
Central Highlands of Ethiopia. J. Agric. Sci. 5, 134–144. https://doi.org/10.5539/jas. Change: Moving Beyond the Myths, vol. 11. pp. 261–269. https://doi.org/10.1016/
v5n2p134. S0959-3780(01)00007-3.
Amsalu, A., de Graaff, J., 2007. Determinants of adoption and continued use of stone Legesse, B.A., Jefferson-Moore, K., Thomas, T., 2018. Impacts of land tenure and property
terraces for soil and water conservation in an Ethiopian highland watershed. Ecol. rights on reforestation intervention in Ethiopia. Land Use Policy 70, 494–499.
Econ. 61, 294–302. https://doi.org/10.1016/j.ecolecon.2006.01.014. https://doi.org/10.1016/j.landusepol.2017.11.018.
Anley, Y., Bogale, A., Haile-Gabriel, A., 2007. Adoption decision and use intensity of soil Li, M., 2012a. Using the propensity score method to estimate causal effects: a review and
and water conservation measures by smallholder subsistence farmers in Dedo practical guide. Organ. Res. Methods 16, 188–226. https://doi.org/10.1177/
District, Western Ethiopia. L. Degrad. Dev. 18, 289–302. https://doi.org/10.1002/ 1094428112447816.
ldr.775. Li, M., 2012b. Using the propensity score method to estimate causal effects: a review and
Badgley, C., Moghtader, J., Quintero, E., Zakem, E., Chappell, M.J., Avilés-Vázquez, K., practical guide. Organ. Res. Methods 16, 188–226. https://doi.org/10.1177/
Samulon, A., Perfecto, I., 2007. Organic agriculture and the global food supply. 1094428112447816.
Renew. Agric. food Syst. 22, 86–108. https://doi.org/10.1017/S1742170507001640. Liang, J., Reynolds, T., Wassie, A., Collins, C., Wubalem, A., 2016. Effects of exotic
Baye, T.G., 2017. Poverty, peasantry and agriculture in Ethiopia. Ann. Agrar. Sci. 15, Eucalyptus spp. Plantations on soil properties in and around sacred natural sites in
420–430. https://doi.org/10.1016/j.aasci.2017.04.002. the northern Ethiopian Highlands. AIMS Agric. Food 1, 175–193. https://doi.org/10.
Becker, S.O., Caliendo, M., 2007. Mhbounds: Sensitivity Analysis for Average Treatment 3934/agrfood.2016.2.175.
Effects (No. 659). DIW Discussion Papers. Berlin. . Nigussie, Z., Tsunekawa, A., Haregeweyn, N., Adgo, E., Nohmi, M., Tsubo, M., Aklog, D.,
Beedy, T.L., Snapp, S.S., Akinnifesi, F.K., Sileshi, G.W., 2010. Impact of Gliricidia sepium Meshesha, D.T., Abele, S., 2017. Factors influencing small-scale farmers’ adoption of
intercropping on soil organic matter fractions in a maize-based cropping system. sustainable land management technologies in north-western Ethiopia. Land Use
Agric. Ecosyst. Environ. 138, 139–146. https://doi.org/10.1016/j.agee.2010.04.008. Policy 67, 57–64. https://doi.org/10.1016/j.landusepol.2017.05.024.
CEFP, 2012. Ecosystem Profile: Eastern Afromontane Biodiversity Hotspot. Nyssen, J., Poesen, J., Moeyersons, J., Deckers, J., Haile, M., Lang, A., 2004. Human
Chepkochei, L.C., 2011. Object-oriented image classification of individual trees using impact on the environment in the Ethiopian and Eritrean highlands - A state of the
erdas imagine objective: case study of wanjohi area, Lake Naivasha Basin. Kenya. art. Earth-Science Rev. 64, 273–320. https://doi.org/10.1016/S0012-8252(03)
Proceedings, Kenya Geotherm. Conf. 00078-3.
Chindi, A., Shunka, E., Solomon, A., Giorgis, G., Seid, E., Tessema, L., 2017. Participatory Oicha, T., Cornelis, W.M., Verplancke, H., Nyssen, J., Govaerts, B., Behailu, M., Haile, M.,
potato seed production : a breakthrough for food security and income generation in Deckers, J., 2010. Short-term effects of conservation agriculture on Vertisols under
the central highlands of Ethiopia. Open Agric. 2, 205–212. https://doi.org/10.1515/ tef (Eragrostis tef (Zucc.) Trotter) in the northern Ethiopian highlands. Soil Tillage
opag-2017-0021. Res. 106, 294–302. https://doi.org/10.1016/j.still.2009.12.004.
Clausen, A., Thomsen, J., Atuhaire, A., Jors, E., 2017. Effect of integrated pest manage- Osman, M., Sauerborn, P., 2001. Soil and water conservation in Ethiopia: experiences and
ment training on Ugandan small-scale farmers. Environ. Health Insights 11. https:// lessons. J. Soils Sediments 1, 117–118.
doi.org/10.1177/1178630217703391. Owenya, M.Z., Mariki, W.L., Kienzle, J., Friedrich, T., Owenya, M.Z., Mariki, W.L.,
Darbyshire, I., Lamb, H., Umer, M., 2003. Forest clearance and regrowth in northern Kienzle, J., Friedrich, T., Kassam, A., 2011. Conservation agriculture (CA) in
Ethiopia during the last 3000 years. Holocene 13, 537–546. https://doi.org/10.1191/ Tanzania: the case of the Mwangaza B CA farmer field school (FFS), Rhotia Village,
0959683603hl644rp. Karatu District. Arusha. Int. J. Agric. Sustain. 9, 145–152. https://doi.org/10.3763/
Davis, K., Nkonya, E., Kato, E., Mekonnen, D.A., Odendo, M., Miiro, R., Nkuba, J., 2012. ijas.2010.0557.
Impact of farmer field schools on agricultural productivity and poverty in East Africa. Pan, D., Kong, F., Zhang, N., Ying, R., 2017. Knowledge training and the change of fer-
World Dev. 40, 402–413. https://doi.org/10.1016/j.worlddev.2011.05.019. tilizer use intensity: evidence from wheat farmers in China. J. Environ. Manage. 197,
de Graaff, J., Amsalu, A., Bodnár, F., Kessler, A., Posthumus, H., Tenge, A., 2008. Factors 130–139. https://doi.org/10.1016/j.jenvman.2017.03.069.
influencing adoption and continued use of long-term soil and water conservation Perfecto, I., Vandermeer, J., 2008. Biodiversity conservation in tropical agroecosystems: a
measures in five developing countries. Appl. Geogr. 28, 271–280. https://doi.org/10. new conservation paradigm. Ann. N. Y. Acad. Sci. 1134, 173–200. https://doi.org/
1016/j.apgeog.2008.05.001. 10.1196/annals.1439.011.
Delelegn, Y.T., Purahong, W., Blazevic, A., Yitaferu, B., Wubet, T., Göransson, H., Pittelkow, C.M., Liang, X., Linquist, B.A., Van Groenigen, L.J., Lee, J., Lundy, M.E., Van
Godbold, D.L., 2017. Changes in land use alter soil quality and aggregate stability in Gestel, N., Six, J., Venterea, R.T., Van Kessel, C., 2015. Productivity limits and po-
the highlands of northern Ethiopia. Sci. Rep. 7, 1–12. https://doi.org/10.1038/ tentials of the principles of conservation agriculture. Nature 517, 365–368. https://
s41598-017-14128-y. doi.org/10.1038/nature13809.
Deressa, T.T., Hassan, R.M., 2009. Economic impact of climate change on crop production Pretty, J., Bharucha, Z.P., 2014. Sustainable intensification in agricultural systems. Ann.
in ethiopia: evidence from cross-section measures. J. Afr. Econ. https://doi.org/10. Bot. 114, 1571–1596. https://doi.org/10.1093/aob/mcu205.
1093/jae/ejp002. Rosenbaum, P.R., Rubin, D.B., 1983. The central role of the propensity score in ob-
Ege, S., 2017. Land tenure insecurity in post-certification Amhara. Ethiopia. Land use servational studies for causal effects. Biometrika 70, 41–55. https://doi.org/10.1093/
policy 64, 56–63. https://doi.org/10.1016/j.landusepol.2017.02.015. biomet/70.1.41.
Eggen, M., Ozdogan, M., Zaitchik, B.F., Simane, B., 2016. Land cover classification in Sachs, J.D., McArthur, J.W., Schmidt-Traub, G., Kruk, M., Bahadur, C., Faye, M., McCord,
complex and fragmented agricultural landscapes of the Ethiopian highlands. Remote G., 2004. Ending Africa’s poverty trap. Brookings Pap. Econ. Act. 1, 117–240.
Sens. 8. https://doi.org/10.3390/rs8121020. Settle, W., Soumare, M., Sarr, M., Garba, Hama M., Poisot, A., 2014. Reducing pesticide
Feder, G., Murgai, R., Quizon, J.B., 2003. Sending Farmers Back to School: the Impact of risks to farming communities: cotton farmer field schools in Mali. Philos. Trans. Biol.
Farmer Field Schools in Indonesia (No. 3022). Washington, D.C.. . Sci. 369. https://doi.org/10.1098/rstb.2012.0277.
Getachew, A., Amare, G., Woldeyesus, S., 2008. Yield potential and land-use efficiency of Shiferaw, B., Aune, J.B., Holden, S.T., 2001. Population pressure and land degradation in
wheat and faba bean mixed intercropping. Agron. Sustain. Dev. 28, 257–263. the Ethiopian highlands: a bio-economic model and endogenous soil degradation. In:
Haile, B., Azzarri, C., Roberts, C., Spielman, D.J., 2017. Targeting, bias, and expected Heenrink, N., van Keulen, H., Kuiper, M. (Eds.), Economic Policy and Sustainable
impact of complex innovations on developing-country agriculture: evidence from Land use - Recent Advances in Quantitative Analysis for Development Countries.
Malawi. Agric. Econ. (United Kingdom) 48, 317–326. https://doi.org/10.1111/agec. Physica-Verlag, Heidelberg, New York, pp. 1–21. https://doi.org/10.1007/978-3-
12336. 642-57558-7.
Haregeweyn, N., Tsunekawa, A., Nyssen, J., Poesen, J., Tsubo, M., Meshesha, D., Schutt, Shiferaw, B., Holden, S., 1999. Soil erosion and smallholders’ conservation in the
B., Adgo, E., Tegegne, F., 2015. Soil erosion and conservation in Ethiopia : a review. Highlands of Ethiopia. World Dev. 27, 739–752.
Prog. Phys. Geogr. 1–25. https://doi.org/10.1177/0309133315598725. Shiferaw, B., Holden, S.T., 1998. Resource degradation and adoption of land conservation
Harris, I., Jones, P.D., Osborn, T.J., Lister, D.H., 2014. Updated high-resolution grids of technologies in the Ethiopian Highlands: a case study in Andit Tid, North Shewa.
monthly climatic observations - the CRU TS3.10 Dataset. Int. J. Climatol. 34, Agric. Econ. 18, 233–247. https://doi.org/10.1016/S0169-5150(98)00036-X.
623–642. https://doi.org/10.1002/joc.3711. Shiferaw, B.A., Okello, J., Reddy, R.V., 2009. Adoption and adaptation of natural resource
Hengl, T., De Jesus, J.M., Heuvelink, G.B.M., Gonzalez, M.R., Kilibarda, M., Blagotić, A., management innovations in smallholder agriculture: reflections on key lessons and
Shangguan, W., Wright, M.N., Geng, X., Bauer-Marschallinger, B., Guevara, M.A., best practices. Environ. Dev. Sustain. 11, 601–619. https://doi.org/10.1007/s10668-
Vargas, R., MacMillan, R.A., Batjes, N.H., Leenaars, J.G.B., Ribeiro, E., Wheeler, I., 007-9132-1.
Mantel, S., Kempen, B., 2017. SoilGrids250m: global gridded soil information based Taffesse, A.S., Dorosh, P., Asrat, S., 2011. Crop Production in. Regional Patterns and
on machine learning. PLoS One. https://doi.org/10.1371/journal.pone.0169748. Trends, Ethiopia.
Holden, S., Shiferaw, B., 2004. Land degradation, drought and food security in a less- Takahashi, R., Todo, Y., 2012. Impact of community-based forest management on forest
favoured area in the Ethiopian highlands: a bio-economic model with market im- protection: evidence from an aid-funded project in Ethiopia. Environ. Manage. 50,
perfections. Agric. Econ. 30, 31–49. https://doi.org/10.1016/j.agecon.2002.09.001. 396–404. https://doi.org/10.1007/s00267-012-9887-5.

11
N.S. Chesterman, et al. Land Use Policy 87 (2019) 104051

Tekle, K., 1999. Land degradation problems and their implications for food shortage in conservation in the northern Ethiopian highlands: impacts on soil fertility and crop
South Wello. Ethiopia. Environ. Manage. 23, 419–427. https://doi.org/10.1007/ yield. Soil Tillage Res. 90, 1–15. https://doi.org/10.1016/j.still.2005.08.004.
s002679900197. World Bank Group, 2015. Ethiopia Poverty Assessment 2014. Washington, D.C.. .
Tolessa, T., Senbeta, F., Kidane, M., 2017. The impact of land use/land cover change on Yimer, F., Ledin, S., Abdelkadir, A., 2007. Changes in soil organic carbon and total ni-
ecosystem services in the central highlands of Ethiopia. Ecosyst. Serv. https://doi. trogen contents in three adjacent land use types in the Bale Mountains, south-eastern
org/10.1016/j.ecoser.2016.11.010. highlands of Ethiopia. For. Ecol. Manage. 242, 337–342. https://doi.org/10.1016/j.
Vancampenhout, K., Nyssen, J., Gebremichael, D., 2006. Stone bunds for soil foreco.2007.01.087.

12

You might also like