**1. Introduction**

Managing the trade-off between short-term provision of human needs in terms of water, food, shelter and fibre for a growing population and maintaining the capacity of the planet to provide these services in the future is a growing global challenge [1–3]. World agriculture is currently both contributing to global environmental change [3,4] and being severely affected by it [5]. It is, therefore, necessary and urgent to transform farming systems from primarily focusing on productivity to instead having social, economic and environmental sustainability at the core of their development [6]. Most food production world-wide is currently unsustainable and needs modification [7–10]. For example, nutrients need to be recycled, external inputs reduced and farming systems diversified [7,10–12]. Simultaneously,

anthropogenic CO<sup>2</sup> emissions are changing the global climate and will continue to do so throughout the 21st century [13]. Crop production and carbon sequestration are ecosystem services with great potential for generating synergies in agricultural landscapes worldwide. Crop production can maintain higher levels of soil carbon through, for example, more perennial crops or returning more crop residues and organic material to fields, and this higher soil carbon level can increase yield [14,15].

The increasing concentration of atmospheric CO<sup>2</sup> is largely attributable to emissions from fossil fuel combustion or land-use change, in the latter case mainly through soil organic carbon (SOC) depletion [13]. Conversion of natural land to agriculture and subsequent soil degradation from, for example, erosion have resulted in average losses of 50–70% of the original SOC pool in agricultural soils [16]. In some areas, this has led to declining yield. There is thus a need for sustainable intensification, where production can increase while the environmental footprint of agriculture decreases [17,18]. This is particularly relevant in developing countries, where most emissions are land-based and where carbon sources can be turned directly into sinks [19]. Sub-Saharan Africa is an area with considerable opportunities for sustainable intensification [17]. Kenya, as a fast-developing country in East Africa, has the potential to take the lead in reaching the Paris Agreement goals on climate change mitigation [20]. In Kenya, 70% of the rural population relies on agriculture for their employment and more than a quarter of the gross domestic product is derived from agriculture [21]. The Kenyan National Climate Change Action Plan 2018–2022 sets goals for agriculture where the main mitigation actions proposed for climate change are limited burning in croplands and more use of conservation tillage and agroforestry [21]. Kenya has the goal of converting 281,000 ha of existing arable and grazing land into agroforestry by 2030 as a climate change mitigation action, and of making climate change-related information and advice an integral part of the Kenyan agricultural advice system [22]. Kenyan smallholders are aware of agroforestry and other sustainable management practices [23] that could increase the soil carbon pool [15,24] and help increase production and income [22,25–28]. The Clean Development Mechanism, which allows countries with carbon emission reduction commitments to implement these reductions in developing countries, can create incentives for smallholders in developing countries to sell carbon sequestered in, for example, agroforestry systems to industrialised countries [29]. Similar incentives can be created within developing countries, through the nationally determined contributions [30]. Engel and Muller [31] identified climate-smart agriculture (e.g., agroforestry) as the most promising practice to be promoted by Payments for Ecosystem Services (PES) among smallholders with limited income. However, the use of carbon finance to incentivise this type of bio-carbon storage is still very low, due to the absence of institutional frameworks, reliable sources of carbon finance and involvement of public and private sector actors [19,32,33]. The number of smallholders needed to achieve an area of land large enough to compensate for project transaction costs also makes carbon finance projects practically unattainable at the current market price for carbon [34,35]. Low carbon prices mean that the incentive for farmers is not the carbon payment, but the benefits arising from emission-reducing farm management. Mbow et al. [36] question whether smallholder farmers can benefit from carbon payments at all and how advisory services can succeed in effectively promoting climate-smart agriculture. More research is needed to identify suitable types of climate-smart agrosystems for different user groups and their results and impacts [36].

The Kenya Agricultural Carbon Project (KACP), a soil and tree carbon project implemented by the non-government organisation (NGO) Vi Agroforestry within the Lake Victoria basin, is targeting small-holder farmers (with <2.5 ha). A majority of the farmers are aiming to be self-subsistent from their farming, largely depending on maize, beans and dairy production, while some are combining the agriculture with off-farm employment or casual jobs. The aim of the KACP is carbon sequestration through the uptake of Sustainable Agricultural Land Management (SALM) practices, enabling smallholder farmers to access the carbon market, as well as increasing yield and productivity and enhancing resilience to climate variability and change [37]. SALM practices include, for example, cover crops and agroforestry to increase biomass production, use of biomass for mulching and composting instead of burning, and avoiding soil erosion through, for example, terracing, reduced tillage and

water harvesting [38] (Table A1). The project provided a dedicated advisory package to promote the use of SALM practices. It also included training in farm enterprise development and village savings and loan associations (VSLA), a microfinance intervention for the accumulation of regular savings within a group and rotating credit opportunities, following definitions by, for example, Bouman [39]. The VSLA system is based on trust among people who know each other and is run by the people themselves in one-year cycles with no outside support except for the initial training. A number of studies have reported on the KACP as a novel type of intervention [40–46]. However, no previous study has examined the effects of different SALM practices on maize yield or the actual maize yield response on farms taking part in the project.

The overall aim of this study was to assess the effects of the KACP on farm productivity and livelihoods during the initial four years (2009–2012) using the uptake of SALM practices, maize yield, food self-sufficiency and savings as indicators. Specific objectives were to (i) assess the level and type of SALM implementation among farmers participating in the KACP over time and compare it with that of neighbouring non-participating farms (control farms); (ii) determine the relationship between SALM practices and maize productivity, and compare the productivity over time and between the KACP farms and control farms; and (iii) quantify the level of food self-sufficiency and savings of the KACP farmers and control farmers. Mechanisms for the spread of knowledge and practices between neighbouring farmers were also discussed.

#### **2. Materials and Methods**

#### *2.1. Characterisation of the Study Areas and Background to the KACP*

The Kenya Agricultural Carbon Project started in 2009 and was implemented in two areas of Kenya, both with around 22,500 ha of potential project area (Figure 1). These were Kisumu (southern part including Kisumu and Siaya counties) and Bungoma (northern part including Bungoma county), with 28 administrative locations (Table A2). The region has a bimodal rainfall pattern with two cropping seasons, normally March–July and August–December (Figure A1). The KACP methodology was developed together with the BioCarbon Fund of the World Bank [37]. In the KACP, smallholder farmers receive carbon credits for soil carbon, which makes it unique since most other community and smallholder carbon projects only support carbon sequestration through tree planting [46]. However, the participating farmers were not given any financial support from the KACP during the study period (2009–2012) and the benefits of using SALM practices were instead emphasised, while the carbon revenue was small and presented to farmers as a co-benefit. Carbon credits were generated and claimed using the approved Verified Carbon Standard methodology, based on background data on soils and climate together with data on the management practices used by the farmers, instead of analysing carbon content in soil samples [47,48]. Carbon revenues were post-paid to farmers in 2014 for the years 2010–2014. Validation and verification were conducted periodically by external teams on emission reductions reported within the KACP [47].

From 2009 to 2012, the advisory system had a fixed number of 28 Vi Agroforestry field advisors, with one advisor in every administrative location, covering approximately 70 km<sup>2</sup> . These field advisors offered advisory services to farmers and facilitated monitoring. They identified registered farmer groups or facilitated group formation, and recruited, informed, trained and contracted smallholder farmers to implement a free choice of SALM practices on their farms. Regular interaction and monitoring of project activities by advisors proved important to avoid misunderstandings and identify risks and challenges in agricultural production as early as possible. No new farmer groups were recruited after 2013, so the number of advisors was reduced to four by 2017, to maintain the information flow. The monitoring data were consolidated in maps and tables at the farmer group level for internal monitoring by farmers. They were also collated at the end of each cropping season and entered in databases and ArcGIS for monitoring and evaluation by Vi Agroforestry.

**Figure 1.** Left: Kenya Agricultural Carbon Project (KACP) areas; the Bungoma sites in Bungoma County (green) and the Kisumu sites in Kisumu and Siaya Counties (purple). Right: The KACP areas **Figure 1.** Left: Kenya Agricultural Carbon Project (KACP) areas; the Bungoma sites in Bungoma County (green) and the Kisumu sites in Kisumu and Siaya Counties (purple). Right: The KACP areas within East Africa.

#### within East Africa. *2.2. Farm Sampling and Data Collection*

From 2009 to 2012, the advisory system had a fixed number of 28 Vi Agroforestry field advisors, with one advisor in every administrative location, covering approximately 70 km2. These field advisors offered advisory services to farmers and facilitated monitoring. They identified registered farmer groups or facilitated group formation, and recruited, informed, trained and contracted smallholder farmers to implement a free choice of SALM practices on their farms. Regular interaction and monitoring of project activities by advisors proved important to avoid misunderstandings and identify risks and challenges in agricultural production as early as possible. No new farmer groups were recruited after 2013, so the number of advisors was reduced to four by 2017, to maintain the information flow. The monitoring data were consolidated in maps and tables at the farmer group The project started with 660 farmer groups involving 10,873 voluntary farmers [49]. By the end of 2012, the KACP had recruited 1555 verified farmer groups with 26,535 farmers, who implemented SALM on 16,490 ha of eligible cropland or grazing land. Among the initial 10,873 project farms, 200 were selected and monitored more closely by the KACP field advisors, in permanent farm monitoring (PFM). To select the 200 PFM farms (100 in each of the project areas), the areas were stratified into agro-ecological zones (AEZ) and the number of farms in each zone was decided in relation to the zone size. A systematic grid of 1.5 <sup>×</sup> 1.5 km<sup>2</sup> for unaligned systematic sampling was applied and the area was divided into clusters formed by the areas between four intersection points of the grid. PFM farms were randomly selected within the clusters according to the agro-ecological stratification [48].

level for internal monitoring by farmers. They were also collated at the end of each cropping season and entered in databases and ArcGIS for monitoring and evaluation by Vi Agroforestry. *2.2. Farm Sampling and Data Collection*  The project started with 660 farmer groups involving 10,873 voluntary farmers [49]. By the end of 2012, the KACP had recruited 1555 verified farmer groups with 26,535 farmers, who implemented SALM on 16,490 ha of eligible cropland or grazing land. Among the initial 10,873 project farms, 200 were selected and monitored more closely by the KACP field advisors, in permanent farm monitoring (PFM). To select the 200 PFM farms (100 in each of the project areas), the areas were stratified into agro-ecological zones (AEZ) and the number of farms in each zone was decided in relation to the zone size. A systematic grid of 1.5 × 1.5 km2 for unaligned systematic sampling was applied and the area was divided into clusters formed by the areas between four intersection points of the grid. PFM farms were randomly selected within the clusters according to the agro-ecological stratification [48]. In this study, the PFM (hereafter called "project") data collected by Vi Agroforestry between 2009 and 2012 were used to determine the uptake rate and correlations of SALM practices to maize yield (Figure 2). In addition, in 2012 a set of 160 control farms (80 in each project area) was selected from the 28 locations of the KACP for the purposes of the present study. When selecting control In this study, the PFM (hereafter called "project") data collected by Vi Agroforestry between 2009 and 2012 were used to determine the uptake rate and correlations of SALM practices to maize yield (Figure 2). In addition, in 2012 a set of 160 control farms (80 in each project area) was selected from the 28 locations of the KACP for the purposes of the present study. When selecting control farms, the project farms were used as reference points and the second farm to the north of the project farm was selected if the owners consented. If that farmer belonged to any group that had worked with Vi Agroforestry, it was skipped and the next farm to the north was asked instead, and so on. Thus, only farms that had not previously worked with Vi Agroforestry were selected as control farms. Data collected from all 360 farms (200 project and 160 control farms) included field size, yield and SALM methods used for all maize fields on the farms during two seasons and four years (2009–2012). However, for the control farms, SALM data were only collected for 2012 and maize yield data for 2009–2011 were obtained retrospectively through farmer recall interviews in 2012 (Figure 2). Data on species, numbers and sizes of farm trees were collected, and months of food self-sufficiency and amount and frequency of savings were ranked by the farmers themselves for 2012 only. All parameters were compared between project and control farms except the effects of individual SALM practices on maize yield, which was done across all farms. The project farmers were monitored for two more years (2013–2014) after this study and the project was then converted to a solely self-monitoring system by all farmers.

farms, the project farms were used as reference points and the second farm to the north of the project farm was selected if the owners consented. If that farmer belonged to any group that had worked with Vi Agroforestry, it was skipped and the next farm to the north was asked instead, and so on. Thus, only farms that had not previously worked with Vi Agroforestry were selected as control farms. monitoring system by all farmers.


SALM methods used for all maize fields on the farms during two seasons and four years (2009–2012). However, for the control farms, SALM data were only collected for 2012 and maize yield data for 2009–2011 were obtained retrospectively through farmer recall interviews in 2012 (Figure 2). Data on species, numbers and sizes of farm trees were collected, and months of food self-sufficiency and amount and frequency of savings were ranked by the farmers themselves for 2012 only. All parameters were compared between project and control farms except the effects of individual SALM practices on maize yield, which was done across all farms. The project farmers were monitored for

**Figure 2.** Available data for project and control farms over the years analysed and the scale of the analysis. All data were compared between project and control farms when possible, except for the effects of SALM and trees on maize yield. **Figure 2.** Available data for project and control farms over the years analysed and the scale of the analysis. All data were compared between project and control farms when possible, except for the effects of SALM and trees on maize yield.

#### *2.3. Data Analysis 2.3. Data Analysis*

All statistical analyses were conducted in R 3.4.2 [50]. Correlations between SALM practices and yield were analysed on the field level and for all available years (2009–2012 for project farms, 2012 for control farms). The effects of each SALM practice on maize yield were examined at the field level. For all nine management practices, a linear mixed effect model (the lme function in the nlme package in R) was created, with maize yield as the response variable, usage (or not) of the nine practices, area, year and season as fixed factors, and farm and pair of farms as random factors. The model included recommended SALM practices, such as (1) no tillage, (2) crop residues for direct mulch, (3) raw manure composting, (4) cover crops, (5) terracing and (6) water harvesting structures, and practices to be avoided due to higher carbon emissions, such as (7) removal of crop residues from fields, (8) applying raw manure to fields and (9) burning of residues. Interactions between different practices were not considered, due to large variations in the degree of implementation of each practice. A model simplification procedure was then used to compare and select the model that best explained the variation in the data. The model comparison was carried out using a step-wise Akaike information criterion (AIC) (allowing both forward and backward comparisons). The best-fit model was run using the lmer function in the lme4 package in R to identify potential significant effects of All statistical analyses were conducted in R 3.4.2 [50]. Correlations between SALM practices and yield were analysed on the field level and for all available years (2009–2012 for project farms, 2012 for control farms). The effects of each SALM practice on maize yield were examined at the field level. For all nine management practices, a linear mixed effect model (the lme function in the nlme package in R) was created, with maize yield as the response variable, usage (or not) of the nine practices, area, year and season as fixed factors, and farm and pair of farms as random factors. The model included recommended SALM practices, such as (1) no tillage, (2) crop residues for direct mulch, (3) raw manure composting, (4) cover crops, (5) terracing and (6) water harvesting structures, and practices to be avoided due to higher carbon emissions, such as (7) removal of crop residues from fields, (8) applying raw manure to fields and (9) burning of residues. Interactions between different practices were not considered, due to large variations in the degree of implementation of each practice. A model simplification procedure was then used to compare and select the model that best explained the variation in the data. The model comparison was carried out using a step-wise Akaike information criterion (AIC) (allowing both forward and backward comparisons). The best-fit model was run using the lmer function in the lme4 package in R to identify potential significant effects of SALM practices.

SALM practices. The effects of agroforestry SALM practices, represented by the number and function/s of farm trees, on maize yield were analysed in a similar way as described above for the SALM practices. The only difference was that the maize yield data used were farm averages from 2012 for each of the two rainy seasons and study areas. For trees, the fixed factors included were area, season and the total number of trees, fodder trees, timber trees and fruit trees. Farm was included as a random factor. The effects of agroforestry SALM practices, represented by the number and function/s of farm trees, on maize yield were analysed in a similar way as described above for the SALM practices. The only difference was that the maize yield data used were farm averages from 2012 for each of the two rainy seasons and study areas. For trees, the fixed factors included were area, season and the total number of trees, fodder trees, timber trees and fruit trees. Farm was included as a random factor.

For comparisons of mean seasonal maize yield per farm over time, a linear mixed effect model (lmer function in lme4 package) was used, with four fixed factors (treatment, area, year, season), which were tested as direct effects and with all two-way interactions included in the model. Only project and control farms forming pairs were included in that analysis, which resulted in 78 pairs of For comparisons of mean seasonal maize yield per farm over time, a linear mixed effect model (lmer function in lme4 package) was used, with four fixed factors (treatment, area, year, season), which were tested as direct effects and with all two-way interactions included in the model. Only project and control farms forming pairs were included in that analysis, which resulted in 78 pairs of farms in Kisumu and 79 in Bungoma. Both individual farms and pairs of farms (project and control) were set as random factors in the model. The paired farms were analysed for four years (2009–2012), and two seasons per year. Contrasts were used to compare the differences in 2009 and 2012 between project and control farms.

The data on food self-sufficiency were categorised into four levels: <6 months food self-sufficiency, 6–7 months, 8–9 months and 10–12 months. A Chi-2 test was used to identify dependencies between food self-sufficiency and KACP participation. The significance level for all analyses was set to *p* < 0.05. project and control farms.

#### **3. Results 3. Results**

#### *3.1. Uptake of SALM and Other Practices 3.1. Uptake of SALM and Other Practices*

Uptake of SALM was studied for all four years for project farms, but only for 2012 for control farms (Figure 2). Project farmers responded well to the advisory services within the KACP and started to implement several of the SALM practices (Figure 3). In 2012, about 60% of the fields were under mulch and terracing, compared with 25% and 40%, respectively, in 2009. Water harvesting increased from around 10% in 2009 to 40% in 2012 and composting of raw manure increased from 50% to 65% (Figure 3). The three most popular SALM practices among project farmers were using crop residues as mulch, composting raw manure before application and terracing fields (Figure 3). Uptake of SALM was studied for all four years for project farms, but only for 2012 for control farms (Figure 2). Project farmers responded well to the advisory services within the KACP and started to implement several of the SALM practices (Figure 3). In 2012, about 60% of the fields were under mulch and terracing, compared with 25% and 40%, respectively, in 2009. Water harvesting increased from around 10% in 2009 to 40% in 2012 and composting of raw manure increased from 50% to 65% (Figure. 3). The three most popular SALM practices among project farmers were using crop residues as mulch, composting raw manure before application and terracing fields (Figure 3).

*Land* **2020**, *9*, x FOR PEER REVIEW 6 of 23

farms in Kisumu and 79 in Bungoma. Both individual farms and pairs of farms (project and control) were set as random factors in the model. The paired farms were analysed for four years (2009–2012), and two seasons per year. Contrasts were used to compare the differences in 2009 and 2012 between

The data on food self-sufficiency were categorised into four levels: <6 months food selfsufficiency, 6–7 months, 8–9 months and 10–12 months. A Chi-2 test was used to identify

**Figure 3.** Diagram showing SALM uptake and implementation in fields in the two areas of Kisumu and Bungoma by project farmers and control farmers between 2009 (project start) and 2012. Data for control groups only available for 2012. Data for project farms in Kisumu in 2009 were lost due to a computer malfunction. N = 183 for control 2012, n = 208 for project 2009, n = 481 for project 2010, n = 591 for project 2011 and n = 495 for project 2012. **Figure 3.** Diagram showing SALM uptake and implementation in fields in the two areas of Kisumu and Bungoma by project farmers and control farmers between 2009 (project start) and 2012. Data for control groups only available for 2012. Data for project farms in Kisumu in 2009 were lost due to a computer malfunction. N = 183 for control 2012, n = 208 for project 2009, n = 481 for project 2010, n = 591 for project 2011 and n = 495 for project 2012.

Some measures never became commonly used, such as no tillage and cover crops. Cover crops increased from 10% in 2009 to 55% in 2011 but decreased again to 10% in 2012. Information on the use of SALM practices on control farms was only available for 2012, at which time crop residues were used for direct mulching on around 50% of control farmers' fields, compared with 25% on project fields in 2009 and 60% in 2012. Raw manure composting was also quite common on control farms, while terracing, use of cover crops and creation of water harvesting structures were rare. The Some measures never became commonly used, such as no tillage and cover crops. Cover crops increased from 10% in 2009 to 55% in 2011 but decreased again to 10% in 2012. Information on the use of SALM practices on control farms was only available for 2012, at which time crop residues were used for direct mulching on around 50% of control farmers' fields, compared with 25% on project fields in 2009 and 60% in 2012. Raw manure composting was also quite common on control farms, while terracing, use of cover crops and creation of water harvesting structures were rare. The measures to be avoided according to the KACP recommendations were reduced on project farms. Removal of residues decreased from 50% to 20%, raw manure application from 30% to 10% and burning of residues from 30% to close to zero. On the control farms, removing residues from fields and burning of residues were still used to a relatively high degree. In general, by 2012, the project farms had on average more of the promoted practices and fewer practices to be avoided according to the KACP recommendations compared with both the start of the project in 2009 and with the control farms.

Tree planting, especially agroforestry, was another promoted practice taken up by farmers. The majority of trees (counted only in 2012) on the farms were young and planted within the project period. On average, farms in Bungoma had a higher number of trees (98 for project farms, 57 for control

farms.

farms) than farms in Kisumu (74 for project farms, 40 for control farms) (Figure 4). Most trees on the farms were timber trees (64% on project farms, 75% on control farms). Apart from having on average more trees (86) per farm compared with control farms (48), project farms also had a larger average proportion (19%) of fodder trees than control farms (4%) (Figure 4). In terms of species, on average control farms had more Eucalyptus spp. and project farms had more N-fixing species like Sesbania sesban, Acacia spp. and Calliandra spp. The most common species overall were Grevillea robusta, Markhamia spp. and Albizia spp. control farms) than farms in Kisumu (74 for project farms, 40 for control farms) (Figure 4). Most trees on the farms were timber trees (64% on project farms, 75% on control farms). Apart from having on average more trees (86) per farm compared with control farms (48), project farms also had a larger average proportion (19%) of fodder trees than control farms (4%) (Figure 4). In terms of species, on average control farms had more Eucalyptus spp. and project farms had more N-fixing species like Sesbania sesban, Acacia spp. and Calliandra spp. The most common species overall were Grevillea robusta, Markhamia spp. and Albizia spp.

period. On average, farms in Bungoma had a higher number of trees (98 for project farms, 57 for

*Land* **2020**, *9*, x FOR PEER REVIEW 7 of 23

measures to be avoided according to the KACP recommendations were reduced on project farms. Removal of residues decreased from 50% to 20%, raw manure application from 30% to 10% and burning of residues from 30% to close to zero. On the control farms, removing residues from fields and burning of residues were still used to a relatively high degree. In general, by 2012, the project farms had on average more of the promoted practices and fewer practices to be avoided according to the KACP recommendations compared with both the start of the project in 2009 and with the control

**Figure 4.** The average number of trees per farm on project and control farms in 2012, divided between fodder, fruit, medicine, timber and other tree types. In total, 65% of all trees were 10 cm or less in diameter at breast height, meaning that they were probably planted within the project period (since 2009). Eucalyptus comprised 4.6% of trees on control farms and 2.7% on project farms. Data from a tree survey in Kisumu (n = 37 control farms, n = 90 project farms) and Bungoma (n = 38 control farms, n = 60 project farms). **Figure 4.** The average number of trees per farm on project and control farms in 2012, divided between fodder, fruit, medicine, timber and other tree types. In total, 65% of all trees were 10 cm or less in diameter at breast height, meaning that they were probably planted within the project period (since 2009). Eucalyptus comprised 4.6% of trees on control farms and 2.7% on project farms. Data from a tree survey in Kisumu (n = 37 control farms, n = 90 project farms) and Bungoma (n = 38 control farms, n = 60 project farms).

#### *3.2. Effects of SALM Practices on Maize Yield 3.2. E*ff*ects of SALM Practices on Maize Yield*

Among the recommended SALM practices, only terracing had a significant (P = 0.0004) positive effect on maize yield. However, the management effects were small compared with the highly significant differences (P < 0.0001) between years, seasons and regions. Terracing was the only practice that was part of the best explanatory model, together with area, season and year (Yield ~ Area + Year + Season + Terracing + (1|Place/Farm)). When analysing effects from trees, first-season maize yield was not affected by any factor other than region, with Bungoma having a higher yield. Second-season maize yield had the best-fit model that included region and the total number of trees per farm (Yield 2 ~ Region + Total trees + (1|Farm)). Second-season maize yield increased with the increasing total number of trees (P = 0.02). Among the recommended SALM practices, only terracing had a significant (*p* = 0.0004) positive effect on maize yield. However, the management effects were small compared with the highly significant differences (*p* < 0.0001) between years, seasons and regions. Terracing was the only practice that was part of the best explanatory model, together with area, season and year (Yield ~ Area + Year + Season + Terracing + (1|Place/Farm)). When analysing effects from trees, first-season maize yield was not affected by any factor other than region, with Bungoma having a higher yield. Second-season maize yield had the best-fit model that included region and the total number of trees per farm (Yield 2 ~ Region + Total trees + (1|Farm)). Second-season maize yield increased with the increasing total number of trees (*p* = 0.02).

#### *3.3. Maize Productivity*

Maize yield varied widely between the areas and was on average 1572–2675 kg ha−<sup>1</sup> in Bungoma, compared with 725–1661 kg ha−<sup>1</sup> in Kisumu, among project farms in the first season of the four years (Table 1). There was also a wide variation in second-season yield, which was 518–1054 kg ha−<sup>1</sup> and 152–678 kg ha−<sup>1</sup> for Bungoma and Kisumu control farms, respectively. Maize productivity analyses showed higher yield (*p* < 0.0001) for project farms than control farms, higher first-season than second-season yield (*p* < 0.0001) and higher yield (*p* < 0.0001) in Bungoma than Kisumu (Table 1, Figure 5). The four study years also differed (*p* < 0.0001) in terms of yield, with mostly increasing trends (Figure 5). Apart from the main effects, there were three significant interactions (Figure 6): (1) yield

differences between project and control farms were larger (*p* < 0.0001) in 2010–2011 than in 2009 or 2012 (Figure 6a); (2) first-season yield increased more (*p* = 0.004) than second-season yield in 2012 (Figure 6b); and (3) there was a larger difference (*p* = 0.005) in the first-season yield compared with the second-season yield between the regions (Figure 6c). Project farms increased their yield mostly in 2010 when the first-season yield declined in control farms. The second-largest increase for project farms was in 2011, while control farms had a one-year lag with their largest increase in 2011 and second-largest in 2012 (although the yield was still lower than on project farms). However, in total, the difference in yield between project and control farms was similar in 2009 and 2012 (*p* = 0.15), and therefore the yield gap between project and control farms did not change significantly during the four initial years of the KACP.

**Table 1.** Maize yield (kg ha−<sup>1</sup> ) for all project and control farms (measured on field level) in Bungoma and Kisumu. Values are mean ± standard deviation for all four years (2009–2012) and both seasons (1 and 2); n = number of farms included in the analysis.


### *3.4. KACP E*ff*ects on Savings and Food Su*ffi*ciency*

Project farmers added to their savings more often on average than control farmers (Figure 7a) and also added larger amounts per occasion (Figure 7b). More than 45% of project farmers added to savings two or more times per month, while the corresponding figure for control farmers was 21%. Project farmers saved to a larger extent (72%) than control farmers (52%) and Kisumu farmers saved on average more than Bungoma farmers. In the VSLAs within the KACP, farmers were able to borrow up to three times their savings to use for investments.

Among Bungoma farmers, 39% had farm inputs as their main expenditure, compared with just 14% for Kisumu farmers, 60% of whom had food as their main cost. In Bungoma, there was a difference between project and control farmers in that 51% of project farmers spent most on education for their children, while 25% of control farmers still had to spend most on food and thereby only 27% had education as their main expenditure. The majority of farmers in both Kisumu and Bungoma had their main source of income from agricultural products.

Only 4% of control farmers in Kisumu (31% in Bungoma) had enough food for 10–12 months, compared with 16% of project farmers in Kisumu (51% in Bungoma). Moreover, 46% of control farmers in Kisumu and 25% in Bungoma had enough food for less than six months, while the corresponding values for project farmers were 19% and 7%, respectively. In general, farmers in Bungoma had more months of food sufficiency than farmers in Kisumu. Chi-2 tests revealed that food sufficiency was significantly higher (*p* < 0.001) overall for project farmers than control farmers.

*Land* **2020**, *9*, x FOR PEER REVIEW 9 of 23

**Figure 5.** First-season (**a**,**b**) and second-season (**c**,**d**) maize yield kg ha−1 on project (blue) and control (green) farms in Kisumu (**a**,**c**) and Bungoma (**b**,**d**), 2009–2012. The lines indicate mean values for each distribution and the dotted line shows the mean for each sub-plot.

**Figure 5.** First-season (**a**,**b**) and second-season (**c**,**d**) maize yield kg ha−1 on project (blue) and control (green) farms in Kisumu (**a**,**c**) and Bungoma (**b**,**d**), 2009–2012.

The lines indicate mean values for each distribution and the dotted line shows the mean for each sub-plot.

*Land* **2020**, *9*, x FOR PEER REVIEW 10 of 23

**Figure 6.** Diagrams showing the three significant interactions identified, between: (**a**) year and treatment (P < 0.0001), (**b**) year and season (P = 0.004) and (**c**) region **Figure 6.** Diagrams showing the three significant interactions identified, between: (**a**) year and treatment (*p* < 0.0001), (**b**) year and season (*p* = 0.004) and (**c**) region and season (*p* = 0.005) on maize yield in kg ha−1 for project and control farms.

and season (P = 0.005) on maize yield in kg ha−1 for project and control farms.

Only 4% of control farmers in Kisumu (31% in Bungoma) had enough food for 10–12 months, compared with 16% of project farmers in Kisumu (51% in Bungoma). Moreover, 46% of control farmers in Kisumu and 25% in Bungoma had enough food for less than six months, while the corresponding values for project farmers were 19% and 7%, respectively. In general, farmers in

**Figure 7.** (**a**) Percentage (*y*-axis) of project and control farm households saving money on different numbers of occasions per month in 2012. (**b**) Percentage of project and control farm households saving different amounts on every saving occasion in 2012. **Figure 7.** (**a**) Percentage (*y*-axis) of project and control farm households saving money on different numbers of occasions per month in 2012. (**b**) Percentage of project and control farm households saving different amounts on every saving occasion in 2012.

### **4. Discussion 4. Discussion**
