Magnitude of wheat yield gaps in Ethiopia
Wheat yields in farmers’ fields were on average 1.9 t ha− 1 for the pooled sample, which corresponds to a yield gap closure of 21% of Yw (Fig. 3). The lowest Ya (less than 1.7 t ha− 1) was recorded in the moist and sub-moist agro-ecological zones (M2, M3, and SM3) while the highest Ya (1.8–2.2 t ha− 1) was recorded in the humid (H2 and H3) and sub-humid agro-ecological zones (SH1 and SH2; Fig. 3A and D). The latter are indeed amongst the most suitable agro-ecological zones for wheat production in Eastern Africa (Negassa et al. 2013; Hodson and White 2007). Ya also varied across administrative zones with a minimum of 1.2 t ha− 1 in South Wollo and North Gonder and a maximum of 2.9 t ha− 1 in West Arsi (Fig. 3B and E). No major differences in Ya were observed between highland mixed and highland perennial farming systems (Fig. 3C and F) while wheat yields were considerably greater in maize mixed farming systems, ca. 3.2 t ha− 1, than in the other farming systems. However, we note the bulk of the sample is classified as highland mixed farming system (Table 2). Finally, Yw varied between 8.3 t ha− 1 in South Gonder (moist sub-afroalpine areas) and in East Showa (sub-moist and sub-humid highlands) up to 10.5 t ha− 1 in West Arsi (humid and sub-humid highlands) and Gurage (sub-humid highlands; Fig. 3A and B).
The actual yield reported in the household survey at national level was 2.1 and 1.7 t ha− 1 in 2009 and 2013, respectively. We acknowledge yield progress may have occurred since then but our analysis is still relevant as it focuses on comparisons between farms and regions. Yet, the mean actual yield in 2009 was slightly greater than that reported in the official statistics in the same year while the opposite was true in 2013 (CSA, 2018). Official statistics indicate actual yields of wheat in Ethiopia of 1.8 and 2.5 t ha− 1 for the years 2009 and 2013, respectively, and clearly show wheat yield progress since the early 2000s (FAOSTAT, 2019). Actual yields for West Arsi in the household survey analysed here (i.e. WIAS) were on average 3.4 and 2.2 t ha− 1 in 2009 and 2013, respectively, values which are in line with the 2.7 t ha− 1 reported in Silva et al. (2019a) for a different household survey conducted in the same zone in 2012.
Wheat yield gaps were mostly attributed to the technology yield gap (> 50% of Yw) but narrowing efficiency and resource yield gaps can still double Ya (Fig. 3). This was true for most agro-ecological zones, administrative zones, or farming systems. The efficiency yield gap was on average 10% of Yw and did not differ much between agro-ecological zones (7.8–10.7% of Yw, Fig. 3D), administrative zones (7.5–11.4% of Yw, Fig. 3E), or farming systems (10.1–10.5% of Yw, Fig 3F). The resource yield gap was on average 15% of Yw and was smallest in the highland agro-ecological zones (SH2, H2, and SM2, 9.3–13.1% of Yw) and greatest in the SM3, M2, M3, and SH1 agro-ecological zones (> 15% of Yw, Fig. 3D). In terms of administrative zones, the resource yield gap was smaller than 10% of Yw in West Arsi, North Gonder, and Gurage and above 20% of Yw in West Showa, South Wollo, and North Wollo (Fig. 3E). The resource yield gap was negligible for maize mixed farming systems and ca. 15% of Yw for highland mixed and highland perennial farming systems (Fig. 3F). These efficiency and resource yield gaps seem to be small when expressed in relation to Yw but we note they are far from insignificant when compared to Ya (Fig. 3). Finally, high seed rates and weed control together with N application rates of 150, 250, and 350 kg N ha− 1 resulted in an average yield gap closure to 50, 60, and 70% of Yw, respectively. For instance, an application of 350 kg N ha− 1 increased wheat yields to ca. 80% of Yw in sub-afroalpine agro-ecological zones (H3 and SM3; Fig. 3D) and to 60–75% of Yw in a number of administrative zones (Fig. 3E) and highland mixed farming systems (Fig. 3F).
In summary, fine-tuning current crop management practices can deliver the additional production needed to reach wheat self-sufficiency without expanding wheat area. However, further narrowing yield gaps towards Yw requires inputs and technologies currently lacking in highest-yielding fields. Technologies currently not used by many farmers include for instance mechanisation of land preparation, planting and harvesting operations, effective control of pests, diseases, and weeds or other nutrients beyond N and P. Efficiency and resource yield gaps as large as current Ya and technology yield gaps as large as 50% of Yw have also been reported in other studies on wheat yield gaps in Ethiopia (Silva et al. 2019a) and Rwanda (Baudron et al. 2019), and on maize yield gaps in Ethiopia (Assefa et al. 2020) and Tanzania (van Dijk et al. 2017).
Drivers of yield variability and gaps at field level
Production frontier and yield variability
The magnitude, sign, and significance of the first-order terms of growth-defining, -limiting, and -reducing factors on wheat yields were consistent between the Cobb-Douglas and translog stochastic models fitted to the pooled sample (Table S1). Regarding the second-order terms, the translog model revealed positive quadratic effects of seed rate, N rate and herbicide use, a negative interaction between seed and N rates, and positive interactions between temperature seasonality and seed rate and hand-weeding, and between available water and herbicide use (Table S1).
Wheat yields decreased with increased growing degrees days and temperature seasonality, after controlling for other factors, and there were no significant differences across varieties and years (Table 3). There was a negative effect of aridity index on wheat yields, a result also found for maize yields in Ethiopia (Assefa et al. 2020). Seed rates had a significant positive effect on wheat yields and increasing the former by 1% resulted in ca. 0.10% increase of the latter. This positive association between plant population and wheat yields was also documented for wheat in Rwanda (Baudron et al. 2019). Crop establishment remains a challenge in smallholder conditions due to manual sowing which leads to large variation in sowing depths and heterogeneous plant populations across the field. Plots where water logging or drought were reported by the farmer yielded 35–45% less than plots where these were not reported, and plots with deeper soils yielded ca. 8% more than plots with medium or shallow soil depths. Frequent ploughing was found to increase wheat yields using the same household survey data (Abro et al. 2018) but in our analysis, this did not translate into significantly greater wheat yields (Table 3). We note the analysis of Abro et al. (2018) focused exclusively on investigating the effect of ploughing frequency on wheat yields while our analysis investigates a broader range of biophysical and management drivers, which overtake the level of yield variation explained by ploughing frequency.
There was a clear yield response to N across models estimated for the pooled sample and for specific administrative regions: on average, wheat yields increased by ca. 0.27% with 1% increase in N applied (Table 3). Earlier studies also identified N fertilisation as a key determinant of wheat yields in Ethiopia (Habte et al. 2014; Tanner et al. 1993) but further research should investigate whether increasing N rates is economically viable for smallholders (cf. van Dijk et al. 2017). High fertiliser prices were identified as an important constraint to increase fertiliser access and use by wheat smallholders in Ethiopia (Anteneh and Asrat 2020). In addition to profitability, smallholders’ decisions to apply fertiliser also depend on the area share of each crop, household wealth, access to rental land, and the level of land fragmentation (Yu and Nin-Pratt 2014). At regional level, the distance from the input distribution to the farmer was found to increase the price of mineral fertilisers as a result of greater transaction and transportation costs (Minten et al. 2013). Overcoming these constraints at farm and regional levels thus remains important to increase access to and use of fertilisers in the country.
Fertile plots yielded 6% and 16% more than medium and poor fertile plots, respectively. No significant yield differences were observed between plots with and without legumes as preceding crop, which is not in agreement with earlier empirical findings (Taa et al. 2004), nor between plots with and without manure application. This may be due to the heterogeneity in manure management and legume productivity and residue management between farms and to the relatively low number of fields with manure use reported (n = 534) and legumes recorded as previous crop (n = 840). Finally, herbicide use was positively associated with wheat yields (but the effects were small), pesticide use translated into 12% greater wheat yields, and disease occurrence reduced wheat yields by ca. 30%. There were no positive significant effects of hand-weeding on wheat yields as this operation might be done after the crop suffers from severe competition from weed as a result of labour shortages or inconvenient working days during more critical periods of the growing season.
The stochastic frontier model with a Cobb-Douglas functional form was fitted to a subset of the data for the administrative zones West Arsi, North Showa, East Gojam, and South Wollo (Table 3). The results obtained for these administrative zones were largely consistent with the results of the national analysis reported above, particularly for seed rate (only non-significant in West Arsi), N application rate (strongly positive in all zones), occurrence of drought (strongly negative in all zones), and occurrence of diseases (strongly negative in all zones). The most notable difference between both national and regional analyses was that the significance of biophysical variables (e.g. growing degrees day, temperature seasonality and aridity index) observed in the former tend to disappear in the latter and wheat yield responses to N were largest in North Showa. This is expected as the pooled sample used in the national analysis exhibits greater variation in biophysical conditions between households compared to the subset used for the regional analyses.
Resource yield gap and yield response to inputs
YHF were 3.4 t ha− 1 in South Wollo, 3.1 t ha− 1 in East Gojam, 4.2 t ha− 1 in North Showa, and 4.5 t ha− 1 in West Arsi (Fig. 4A and Table 2). YAF and YLF were higher in West Arsi (2.2 and 0.8 t ha− 1), intermediate in East Gojam (1.6 and 0.8 t ha− 1), and North Showa (1.5 and 0.4 t ha− 1) and lower in South Wollo (1.2 and 0.4 t ha− 1).
YHF were associated with significantly greater seed and N application rates compared to YAF and/or YLF across all four administrative zones (Fig. 4B and C). Seed rates in highest-yielding fields were ca. 250 kg ha− 1 in North Showa, East Gojam, and South Wollo, which was significantly greater than the average 180 kg ha− 1 used in average- and lowest-yielding fields. The variation in seed rates between field classes was smaller (and not significant) in West Arsi compared to other administrative zones: ca. 220 and 190 kg ha− 1 in highest- and lowest-yielding fields, respectively. N application rates in highest-yielding fields were ca. 90 kg N ha− 1 in North Showa, East Gojam, and South Wollo, which was significantly greater than the ca. 60 kg N ha− 1 used in average-yielding fields in North Showa and East Gojam, the ca. 45 kg N ha− 1 used in South Wollo, and the ca. 30 kg N ha− 1 (60 kg N ha− 1) observed across the lowest-yielding fields in North Showa and South Wollo (East Gojam).
Labour use for land preparation, sowing, hand-weeding and harvesting was significantly greater for YHF than for YAF and YLF in all administrative zones except West Arsi (Fig. 4D and E). Significant differences in herbicide use across groups were only observed in West Arsi and North Showa (Fig. 4F). As an example, highest-yielding fields were associated with a total labour use of ca. 140, 120, and 100 person-day ha− 1 in South Wollo, North Showa, and East Gojam, respectively, while labour use ranged between 60–80 person-day ha− 1 in the lowest-yielding fields of these zones. Considerably more labour was used in North Showa, East Gojam, and South Wollo than in West Arsi and there was an inverse relationship between labour use for hand-weeding and herbicide use (Fig. 4E and F). This is best seen in West Arsi where herbicide use was greatest (ca. 0.8 L ha− 1) and labour for hand-weeding was lowest (ca. 12 person-day ha− 1), data which validate those of an independent household survey conducted in 2012 in the same region and analysed by Silva et al. (2019a).
Technology yield gap and increased amounts of inputs
The simulated Yw in Arsi administrative zone was 8.6 t ha− 1 in 2009 and 9.7 t ha− 1 in 2013, which was considerably greater than the values observed for YHF during the same years (Figs. S5A and S5B). The variety trials described by Bezabih et al. (2018) were conducted at Kulumsa Agricultural Research Center (KARC), Arsi administrative zone, in 2016 and 2017 and simulated Yw was 9.3 and 9.1 t ha− 1 in these years, respectively. The yields observed in these trials ranged between 4.9 and 7.7 t ha− 1 in 2016 and between 4.4 and 7.0− 1 in 2017 (Figs. S5C and S5D). Despite differences in varieties used in the highest-yielding fields and the variety trials (data not shown), most of the varieties cultivated in highest-yielding fields were improved genotypes bred at KARC with parental material from CIMMYT (Fig. S4). This means farmers use improved varieties that can reach up to ca. 80% of Yw on-station and, hence, that the technology yield gap is likely caused by other factors than lack of improved varieties.
The low amount of inputs (particularly seeds and fertilisers) used in highest-yielding fields compared to what is needed to reach Yw (Fig. 4C and Section 2.3.3) and the lack of certain inputs and technologies in these fields are the most likely drivers of the technology yield gap of wheat across Ethiopia. The former is reflected by the difference between the feasible yield (Yf) and YHF, and it is depicted in Fig. 3 as the additional resource yield gap for increasing amounts of N applied (150, 250, and 350 kg N ha− 1). For instance, combining high seed rates with intensive weeding practices and 150 kg N ha− 1 can increase wheat yields up to an average Yf of 4.6 t ha− 1. This corresponds to a technology yield gap of ca. 47% of Yw. Applying N rates of 250 and 350 kg N ha− 1 result in average Yf of 5.3 and 5.8 t ha− 1 which reduces the technology yield gap to about 39 and 33% of Yw, respectively. We note the aforementioned N application rates are way above those currently observed in highest-yielding fields (Fig. 4C) but these high N rates could be reduced if efficient N management practices are adopted cf. (Assefa et al. 2020; ten Berge et al. 2019). Another point of concern is that most wheat in Ethiopia is currently cultivated in acid soils where yield responses to N are not always clear and hence, increasing the N rates further aggravates soil acidity and lowers wheat yield (Regassa and Agegnehu 2011). This clearly indicates that the scope to increase fertiliser rates is context specific and needs to be integrated with other soil management practices and excellent agronomy. Other factors explaining this yield gap may include poor crop establishment and poor weed control, which currently rely heavily on draught power and manual labour, and poor pest and disease control (as partly shown in Table 3). It is also important to consider that row planting improves radiation interception under high seed rates (high plant populations) as compared to the current farmer practice of broadcasting (Alemu et al. 2014).
Farming systems analysis across different zones
Crop diversity at farm level
The total cultivated land area per farm was on average 2 ha in West Arsi and North Showa and 1.6 and 1.4 ha in East Gojam and South Wollo, respectively (Figs. 5 and S7). This land was allocated differently to different crops in different administrative zones. The share of wheat in the total cultivated land was high in South Wollo and West Arsi, on average ca. 45%, and low in North Showa (35%) and East Gojam (25%). In West Arsi, households allocated 47% and 6% of their cultivated land to other cereals (mostly barley) and to legumes (mostly faba bean), respectively (Fig. 5A). In North Showa, the share of other cereals (mostly barley and red tef) and legumes (mostly faba bean) of the total cultivated land was ca. 37% and 27%, respectively (Fig. 5B). In East Gojam, around 60% of the cultivated land was allocated to other cereals (mostly red and white tef) and only ca. 10% was cultivated with legumes (Fig. 5C). In South Wollo, both other cereals and legumes were cultivated on ca. 25% of the total cultivated land (Fig. 5D). This indicates farms in North Showa and East Gojam are more diverse regarding the crop types cultivated than farms in West Arsi and South Wollo (Fig. 4D and E). The different crop types are known to compete for labour during key periods of the growing season (Silva et al. 2019a), but we were not able to find clear substitution or competition for land and labour between wheat and other crops possibly due to a lack of information on the timing of different operations (Fig. S6). Further research is thus needed to clarify the importance of wheat as a source of income, and priority for investment, in the more diversified farming systems.
Availability of land, labour, and capital
Oxen ownership was associated with slightly greater wheat yields in West Arsi, North Showa, and East Gojam but the effects were only significant in East Gojam (Fig. 6A). In addition, households with more oxen pairs tended to cultivate larger wheat areas than households with few oxen pairs (Fig. 6B). This was particularly true in West Arsi and North Showa, where land is more ‘abundant’ (Fig. 5A and B), and not as much and significantly in East Gojam and South Wollo, where land is constrained (Fig. 5C and D). No significant differences in total labour use for wheat were observed for different levels of oxen ownership in either zone (Fig. 6C); hence, oxen ownership did not translate into labour savings per unit land and into substitution of manual labour by draught power. The economic value of farm assets increased on average with increasing oxen ownership, which was particularly clear in West Arsi and North Showa (Fig. 6D). No major significant differences in input use for wheat were observed across different levels of oxen ownership in either zone (Fig. S8).
In summary, oxen ownership was a proxy for draught power and capital availability and was associated with larger wheat area, particularly in the administrative zones with largest cultivated area per household (i.e. West Arsi and North Showa; Silva et al. 2019a). Hence, access to draught power and capital translates into increases in wheat production through expansion of cultivated land and not so much through intensification of wheat production via yield gap closure (Fig. 6A and B).
Comparisons across administrative zones
The four administrative zones analysed in greater depth in this study capture differences in the level of intensification of wheat production (Fig. 4) and in farming systems regarding the crop area shares and oxen ownership (Fig. S1). Wheat yields were greatest in West Arsi, intermediate in North Showa and East Gojam, and smallest in South Wollo, while the opposite was true for labour use for wheat (both total and hand-weeding; Fig. 4). West Arsi is distinct from the other zones mostly because herbicides are widely used, substituting labour for hand-weeding and possibly other inputs (e.g. N) in the short term (Fig. 4C and E). We also note that the cultivated area per farm is greatest in West Arsi and smallest in South Wollo, which may explain the use of herbicides in the former and the heavy reliance on human labour in the latter (Fig. 5). Finally, there was a positive relationship between the number of pairs of oxen (a proxy for capital availability) and the wheat area cultivated per farm in West Arsi and North Showa, the zones where cultivated land per farm was greatest, while no relationship was observed between the number of pairs of oxen and the input use for wheat (Fig. 6). This means that increases in wheat production are mostly obtained through increases in cultivated areas rather than through yield gap closure and that households with more capital do not necessarily use more inputs for wheat. These results suggest that smallholders do not have proper access to inputs because these are either too expensive or unavailable when needed, which can be pointed as the main challenge for intensification of wheat production and the achievement of wheat self-sufficiency in Ethiopia without expansion of wheat area.