1 Introduction

With a need to meet the food demands of the expected 9 billion population by 2050, soil fertility management is a critical component as it directly impacts crop productivity and environmental sustainability [44]. However, Havlin and Heinger, [11], highlighted that over 50% of the global soils are degraded. Moreover, 45% of the countries with highly degraded soils are in Sub-Saharan Africa (SSA) [35]. According to Amare et al. [3], SSA loses about 22 kg of nitrogen, 2.5 kg of phosphorus, and 15 kg of potassium per hectare of cultivated land per year.

In Uganda specifically, soils are known to be highly degraded with low levels of key nutrients like Phosphorus, Nitrogen, and Potassium while soil pH is generally acidic [28]. This has majorly been attributed to the rush to use chemical fertilizers without correct soil fertility knowledge [30]. Research has revealed that blind use of chemical fertilizers leads to soil degradation, soil acidification, and crust disturbance which lowers organic matter and humus content, stunts plant growth, imbalances soil pH, increases the incidence of pests, and releases greenhouse gases [36]. As a result, the focus has shifted from just the use of fertilizers for improved soil fertility to the efficient use of chemical fertilizers to mitigate their negative impact [14].

Unfortunately, soil testing that is recommended for efficient use of chemical fertilizers has been limitedly adopted, especially among the smallholder farmers who are even the majority in Sub-Saharan Africa [36]. This is attributed to the high cost of soil testing and limited access to soil testing services [6]. Thus, farmers in Sub-Saharan Africa have remained reliant solely on indigenous soil fertility knowledge [14, 15]. In a rush to use chemical fertilizers, soil colour, crop vigour, weeds, previous yield, and sometimes soil water availability, erosion history, presence of soil organisms, and organic matter content are used to assess soil fertility [42, 46]. Yet, these indicators are asserted to be inefficient [15, 34]. However, with the increasing privatization of advisory services in developing countries [20], more research is required on indigenous soil fertility indicators that seem to be simple and affordable, especially for resource-constrained farmers [23].

Although most of the studies on indigenous and scientific soil fertility assessment seem to all agree that indigenous approaches are inefficient [15, 16, 34], there is still limited or no systematic analysis information on the individual performance of the indigenous approaches regarding soil fertility assessment. Therefore, it was imperative to investigate which indigenous approach is better performing amongst the indigenous approaches in comparison to scientific soil testing [17]. But also, under what conditions does it perform better. Thus, we needed to also understand the drivers (correlates) of soil fertility misclassification amongst all the indigenous approaches used by smallholder farmers.

The contribution of this study is to (1) assess the accuracy of indigenous soil fertility assessment approaches commonly used by farmers in Uganda. (2) Characterize farmers who are more likely to misclassify their soil fertility based on indigenous soil fertility indicators and (3) determine the correlates of misclassification associated with the indigenous approaches employed by farmers. To achieve these, we used descriptive statistics and the Probit model with robust standard errors.

2 Materials and methods

2.1 Data

The data used in this study was collected in 2021 from the Agriculture Cluster Development Project (ACDP) beneficiaries in Clusters: 1 (Rakai), 2 (Namutumba), 3 (Butebo, Pallisa and Tororo), 4 (Mbale), 8 (Kamwenge), 9 (Kyegegwa), 11 (Ntungamo). These clusters were selected because they had the highest number of farmers who redeemed fertilizers from the project. The sampling design consisted of probabilistic sampling techniques i.e. stratified sampling and systematic sampling. The project beneficiary list was obtained from the e-voucher system database and then stratified by districts. Using a systematic sampling scale of 2.5 km distance, a sample of 10 farmers was picked from each sub-county but allowance was given where the distance between the sampling units exceeded the sampling scale. Brewer, [5] urged that the allowance be given in consideration of the asymptotic systematic sampling with unequal probabilities as the samplings units tend to zero.

For the soil samples, these were taken at a soil depth of 0–30 cm using the zig-zag sampling method. Samples were taken with a soil sampling auger/corer which enabled ‘cores’ of soil to be taken from below the surface. At least 12 individual cores were taken from each area to make up the composite soil sample for analysis. The depth of penetration for soil cores depended on the nature of the terrain and the nutrients under test. An imaginary ‘W’ covering the area being sampled was employed and cores were not taken close to hedgerows, under trees, or adjacent to buildings. To provide the sample for analysis, individual cores were thoroughly mixed in a basin and foreign materials were removed by hand. The composite samples were then transferred to a Soil Sample bag (PT 301) and labelled for reference. The soil samples were delivered by the sampling agent from the sub-county to the District Focal Person for verification before transportation to the testing laboratory.

The sample labels were attached immediately after soil sample collection and included information such as sample number, date of sampling, farmer’s name, and farmer’s sub-county. The other information such as current land use, sampling depth, total acreage, GPS coordinate of the sample, and the person carrying out the sampling was collected using Kobo Collect. A photo of the sample label was taken and immediately sent to the laboratory within three hours of collecting the sample to enable the laboratory staff to prepare adequately to receive the samples. The authentication of the soil sample information was enhanced using QR codes as an additional feature of sample labelling, laboratory testing, and linking of the test results to the farmer/garden data.

The soil testing for pH, Conductivity, Nitrogen, phosphorus, Potassium, Calcium, and Magnesium was carried out using a variety of techniques and equipment. The Palintest SKW 500 Complete Soil Kit and other accessories as listed in the different methods of testing were used to produce soil test results from the Soil Test 10 photometer. The received soil samples were air-dried on a tray or plastic sheet. Once dried, the samples were stored in a soil sample bag for testing at a more convenient date or laboratory. The soil was passed through a 2 mm sieve before testing at the laboratory. Soil nutrient testing & garden profiling for ACDP beneficiaries included mixing a fixed volume of soil with deionized water in a known ratio with a defined addition of specific extract reagents to promote the release of nutrients from the soil sample. The extraction methods used were identical for some parameters and in such a case, the same extract solution produced was used for multiple parameter testing. Specific sample preparation steps were used for some parameters and the different steps are summarized in Table 1.

Table 1 Extraction methods that were used for soil testing
Full size table

2.2 Modelling soil fertility index and misclassification in soil fertility assessment

Although soil fertility is defined by physical, chemical, and biological indicators, soil pH with its enormous influence on biochemical processes can also be used as a master soil variable for soil fertility assessment [27]. Following studies by Mukherjee and Lal [22], Nehrani. [24] and Zhou et al. [45], this study employed the Principal Component Analysis (PCA) to estimate the soil fertility index using the Integrated soil Quality Index (IQI) (Table 2). According to Zhou et al. [45], the soil fertility index can be generated using either the Nemoro Quality Index (NQI) model or the Integrated Quality Index (IQI) model. However, IQI is most suitable for soil taken at the depth of 0–10 cm or 10–20 cm while NQI is suitable for soil taken at the depth of 20–30 cm. Moreover, Derakhshan et al. [8] also used both NQI and IQI models but IQI produced more plausible results.

Table 2 Outcomes from the principal component analysis (PCA)
Full size table

From the PCA outcomes, soil pH, Phosphorus, Nitrogen, and potassium were selected parameters for the minimum dataset, and their observations were transformed using linear scoring functions (more is better and optimum) see (Table 2).

The sub-weight value of different soil parameters was summed up to 1 under each soil functional property. Once the selected observation was transformed into numerical scores ranging between 0 and 1, a weighted additive approach was used to integrate them into a soil fertility index. From the PCA outcomes, each principal component explained the amount of variation in the dataset which was divided with the maximum total variation of the selected principal components in the minimum datasets to get a weighted value under a particular principal component.

Therefore, the weighted additive was captured as an equation

$$SQ{I}_{PCA}= \left(\sum {W}_{i}-\text{min}\left(\sum {W}_{i}\right)\right)/\left(\text{max}\left(\sum {W}_{i}\right)-\text{min}\left(\sum {W}_{i}\right)\right)$$
(1)

where

$${\text{W}}_{\text{i }}=\left(\frac{\text{individual component explained variation}}{\text{total cummulative explained variation}}\right)\times \text{corresponding principal component}$$
(2)

From the weighted PCA results, the soil fertility index (SFI) was generated and categorized as poor (< 0.3), moderate (0.3–0.6), and good (> 0.6). Cross-tabulations were used to determine false positives and false negatives after comparing the soil fertility index from the PCA and indigenous farmer ratings to establish the level of misclassification based on [9, 21].

2.3 Empirical models

2.3.1 Probit model

To examine the correlates of soil fertility misclassification, the observed outcome variable y equals 1 if the farmer misclassified his soil fertility by use of the indigenous indicator and equals 0 otherwise. The chance of misclassifying the soil fertility denoted by p can be estimated using either a Probit or Logit model or the linear probability model (LPM) considering the discrete nature of the outcome variable. However, the probabilities predicted by the LPM model can at times be negative, violating the rule that probabilities can only lie between 0 and 1 [2]. This weakness of the LPM makes the logit or Probit models the best candidates for modelling binary outcomes. However, whereas generalized least square models may solve the problem of heteroscedasticity, the problem of estimating parameters of threshold decision models remains unresolved when truncating dependent variables through Logit analysis [12]. Thus, in the current study, descriptive statistics and the Probit model were employed using Stata 16. A Probit model was selected because of its ability to generate bounded probability estimates for each observation [37] as well as the assumption of a standard normal distribution [43].

Using the Probit model, the probability P of a farmer misclassifying soil fertility is expressed as a function of underlying predictor variables represented by a vector X. Since Y is the observed binary outcome that a farmer misclassifies soil fertility, underlying continuous unobservable or latent variable \({y}^{*}\) can be expressed as the following single model:

$$Y^{ * } = x^{\prime}\beta + \mu$$

Although \({y}^{*}\) is not observed, we can observe that;

$$Y = \left\{ {\begin{array}{*{20}c} {1,} & {Y^{*} > 0} \\ {0,} & {Y^{*} \le 0} \\ \end{array} } \right.$$

Therefore,

$$\left( {Y_{i} = 1{|}X} \right) = p \left( {X^{\prime}\beta + \mu > 0} \right) = F\left( {X^{\prime}\beta } \right)$$
(3)

2.3.2 Multivariate probit model

In a case where a farmer jointly uses all the indigenous approaches to ascertain their soil fertility, the dependent variables are not mutually exclusive, the multivariate Probit model becomes more appropriate [38]. Moreover, several studies have already used the multivariate Probit model to assess the determinants of interdependent outcome variables [26, 29, 31]. The advantage of this model is the capacity to model simultaneously the misclassifications from different indigenous soil assessment approaches and investigate all the possible drivers of misclassification [4] Assuming the error terms are mutually exclusive, [4] highlight that the multivariate model can now be specified as;

$${{{\varvec{Y}}}^{\boldsymbol{*}}}_{{\varvec{i}}{\varvec{j}}}={{{\varvec{X}}}^{\boldsymbol{^{\prime}}}}_{{\varvec{i}}{\varvec{j}}}{{\varvec{\beta}}}_{{\varvec{j}}}+\boldsymbol{ }{\mathcal{E}}_{{\varvec{i}}{\varvec{j}}}$$
(4)

where;

\(i\) is the farmer identifier.

\({{Y}^{*}}_{ij}\)(j = 1…4) represents an unobserved latent variable of the misclassification from previous yield, vegetation cover, soil colour and weeds indicators.

\({{\varvec{X}}}_{{\varvec{i}}{\varvec{j}}}\) is a \(1\times k\) vector of explanatory variables that influence the misclassification from indigenous soil fertility assessment approaches.

\({{\varvec{\beta}}}_{{\varvec{j}}}\) is \(k\times 1\) vector of unknown model parameters to be estimated.

\({\mathcal{E}}_{{\varvec{i}}{\varvec{j}}}\) are normally distributed multivariate error terms.

Each \({{Y}^{*}}_{ij}\) is a binary variable and therefore the equation will be a system of m equations as shown below;

$$\left\{ {\begin{array}{*{20}c} {Y_{11}^{ * } = X^{\prime}_{11} \beta_{1} + \varepsilon_{1} ,Y_{1} \,\,if\,\,Y^{ * }_{1} > 0,Y_{1} = 0\,\,otherwise} \\ {\begin{array}{*{20}c} {Y_{22}^{ * } = X^{\prime}_{22} \beta_{2} + \varepsilon_{2} ,Y_{2} \,\,if\,\,Y^{ * }_{2} > 0,Y_{2} = 0\,\,otherwise} \\ {\begin{array}{*{20}c} {Y_{33}^{ * } = X^{\prime}_{33} \beta_{3} + \varepsilon_{3} ,Y_{3} \,\,if\,\,Y^{ * }_{3} > 0,Y_{3} = 0\,\,otherwise} \\ {Y_{44}^{ * } = X^{\prime}_{44} \beta_{44} + \varepsilon_{4} ,Y_{4} \,\,if\,\,Y^{ * }_{4} > 0,Y_{4} = 0\,\,otherwise} \\ \end{array} } \\ \end{array} } \\ \end{array} } \right.$$

The error terms across the four misclassifications are assumed to follow a multivariate normal distribution with a mean vector equal to zero and a covariate R matrix with diagonal elements equal to one and correlations\(\rho_{jk} = \rho_{kj}\) as off-diagonal elements (Green, 2002).

Diagnostic tests to detect the presence of multicollinearity were performed using correlation coefficients and variance inflation factor (VIF) for all the variables in the model. Marginal probabilities were computed to predict the effect of change of a predictor variable on the probabilities of farmers misclassifying their soil fertility.

2.4 Diagnostic tests

2.4.1 Correlation between the study variables

The association of the explanatory variables was examined using the Pearson correlation coefficients as shown in (Table 3). The results showed that plot size was positively and significantly associated with soil fertility misclassification while age had a significant negative correlation with misclassification. There was also a significant negative correlation between age and growing crops on sloppy land. The results suggest that as farmers grow old, they are less likely to grow crops on a sloppy land surface. Plot size was positively correlated with group membership and cultivation on owned land. These results imply that the use of large plot sizes for single-crop production is associated with group membership and the use of owned land. Group membership was also significant and positively correlated with the plot's location and land ownership. These results suggest that farmers who belong to the farmer groups are more likely to plant the project-supported crops on owned land which is located on the sloppy land surface.

Table 3 Pearson correlation coefficients showing the relationship among the study variables
Full size table

However, the mean Variance Inflation Factor (VIF) of the explanatory variables was found to be 7.04, which was below the threshold VIF of 10. This implies that there was no presence of multicollinearity among the variables. To check the level of agreement between the indigenous soil fertility assessment approaches and scientific rating, Kendall’s W was used. The results showed a low level of agreement between the indigenous approaches and the scientific rating (W = 0.310, p < 0.01). Implies that the Indigenous soil fertility ratings agree with scientific ratings to a small extent.

3 Results

3.1 Soil fertility assessment using indigenous & scientific approaches

We examined farmers’ soil fertility perceptions based on the four commonly used indigenous soil fertility assessment approaches, which included weeds, vegetation cover, soil colour, and previous yield indicators. We also scientifically tested the soil samples from the same farmers using the Palintest. The weeds indicator considered the common weeds farmers believe to predict soil fertility status such as Amaranthus spinosus, Bidens pilosa, Commelina benghalensis, Striga, etc. The vegetation cover indicator considered the vigour or performance of plant growth in the sampled plot. For the soil colour indicator, the focus was on the soil appearance or soil colour, where farmers often perceive dark soils to be fertile and lighter soils to be infertile while the yield indicator referred to the previous yields that were obtained in the plot of interest.

Based on the weeds’ indicator, 9.98% (46) of the farmers reported poor soils, 81.78% (377) farmers reported moderate soils and 8.24% (38) reported their soils as good. Using vegetation indicator, 10.85% (50) farmers reported poor soils, 80.69% (372) reported moderate soils and 8.46% (39) reported good soils. Considering the previous yield indicators, 21.48% (99) farmers reported poor soils, 71.37% (329) reported moderate soils, and 7.16% (33) reported good soils. Using the soil colour indicator, 13.23% (61) farmers reported poor soils, 80.91% (373) moderate and 5.86% (27) farmers reported good soils. However, the scientific test results, indicated that 67.25% (310) soil samples were poor, 28.42% (131) were moderate and only 4.34% (20) were of good soil fertility as shown in Fig. 1. This implies that the indigenous soil fertility assessment approaches were more likely to correctly predict the fertility of good soils compared to that of poor soils. Besides, most farmers perceived their soils as moderate in fertility, yet they were actually of poor fertility (Fig. 1).

Fig. 1
figure 1

Soil fertility ratings from different assessment approaches

Full size image

3.2 The extent of misclassification associated with indigenous soil fertility assessments

Using the scientific approach as a gold standard, we estimated the extent of misclassification of soil fertility from four commonly used indigenous soil fertility indicators (Fig. 2). The results reveal that 71% of the farmers misclassified their soil fertility using the vegetation cover indicator. This implies that vegetation cover was the most misinterpreted indicator among the four commonly used Indigenous soil fertility indicators. On the hand, 67% of the farmers misclassified their soil fertility using the weeds indicator, 66% misclassified using soil colour, and 61% misclassified using the previous yield indicator. Based on these results, the previous yields indicator presents a lower extent of misclassification compared to the other indigenous soil fertility assessments.

Fig. 2
figure 2

Extent of soil fertility misclassification from indigenous approaches

Full size image

3.3 Misclassification differences between the different indigenous approaches

After determining the different misclassifications levels, we assess the statistical significance of the differences within the four indigenous approaches. The results show that between the previous yield and vegetation cover, the difference is statistically significant and negative. Implying that the previous yield indicator had a lower misclassification level compared to vegetation cover. Comparing the previous yield and the weeds indicator showed that the previous yield had a lower misclassification level compared to the weeds’ indicator.

For the previous yield and soil colour, the results show that the previous yield performs better than the soil colour indicator. This implies that compared to all the other three indigenous approaches, the previous yield indicator had the lowest misclassification rates. However, there was no significant difference between the weeds’ indicator and the soil colour (Table 4).

Table 4 Misclassification differences between indigenous approaches
Full size table

3.4 Characterization of farmers based on soil fertility misclassification

The general characteristics of the sampled farmers by their soil fertility classification status are presented in (Table 5). The results show that the older farmers were more likely to correctly predict their soil fertility using the indigenous approaches compared to younger farmers. These results suggest that older farmers were more experienced and knowledgeable about their soil fertility when using indigenous approaches compared to younger farmers. The results also show that the farmers in the farmer groups were more likely to misclassify their soil fertility compared to non-farmer groups’ members, especially using the previous yield indicator. Regarding the land topography, the proportion of farmers whose farms were in sloppy land who correctly predicted their soil fertility was significantly higher compared to those whose farms were located in flat land across all the four Indigenous indicators. For the case of land ownership, the proportion of farmers who misclassified the soil fertility of the owned plot using the vegetation cover indicator was significantly higher than those who misclassified the soil fertility of the hired plots. This is likely because when individuals hire land, they tend to be more critical about the land quality compared to the landowners. No evidence was found to suggest that education, access to extension services, and plot size significantly influenced soil fertility misclassification at a 5% significance level.

Table 5 Characterization of farmers by soil fertility classification
Full size table

3.5 Correlates of soil fertility misclassification (probit model)

The correlates of soil fertility misclassification based on the four indigenous approaches were estimated using independent Probit Models (Table 6). The likelihood ratio tests showed that the estimated models were statistically significant (\({\upchi }^{2}\) (13) = 44.89, P < 0.01, Average Pseudo \({\text{R}}^{2}\)= 0.0795). Therefore, the null hypothesis that the explanatory variables are jointly zero can be rejected at a 5% level of significance. The results revealed that only the plot size positively and significantly influenced soil fertility misclassification from all the four Indigenous indicators assessed. The land tenure system (owning land) also positively and significantly influenced soil fertility misclassification only when the vegetation cover indicator was used.

Table 6 Correlates of soil fertility misclassification (Probit Model)
Full size table

From the marginal effect estimations, an increase in the plot size by an acre increases the likelihood of soil fertility misclassification when using weeds, vegetation cover, previous yield, and soil colour indicators by 26.7%, 15.8%, 25%, and 26.7%, respectively. However, the vegetation cover presents a slightly lower likelihood of misclassification than the other indigenous approaches even with large plot sizes. On the contrary, the land tenure system used (being a land owner), increased the probability of soil fertility misclassification by 24.4% when using the vegetation cover compared to those farmers who hired land.

3.6 Correlates of soil fertility misclassification (multivariate probit model)

Whereas farmers sometimes use the indigenous approaches independently, there are possible of using all the indigenous approaches to determine the soil fertility. In this section, we explore the correlates of soil fertility misclassification using the multivariate Probit model (Table 7). This is due to significant correlation between the indigenous soil fertility assessment approaches used by the farmers. The Likelihood ratio test of \({\rho }_{ki}\) rejects the null hypothesis of correlation of the error terms hence justifying our use of the multivariate probit model. Some of the results herein are consistent with the independent probit models results shown above.

Table 7 Correlates of soil fertility misclassification (Multivariate probit model)
Full size table

The results reveal that the negative and statistically significant coefficient of age of the farmer reduces the likelihood of soil fertility misclassification when farmer uses the weeds indicators. Implying that elderly farmers are more accurate in assessing soil fertility using weeds compared to any other indigenous approach.

The positive and significant coefficient of access to extension services especially having training on sustainable land management practices increased likelihood of misclassifying when using vegetation cover. This implies that farmers who have access to extension services are more likely to misclassify the fertility of their soils based on the vegetation cover.

The larger the plot size, the higher the likelihood of misclassifying soil fertility across all the indigenous approaches used. The likelihood of misclassifying based on previous yields is higher when the plot sizes are larger compared to other indigenous approaches. For the land tenure system, ownership of land positively and significantly increases the likelihood of misclassifying the soil fertility based on vegetation cover compared to those individuals who rent land.

4 Discussion

In this study, we examine the accuracy of indigenous soil fertility approaches compared to scientific testing and highlight the correlates of misclassification. Our findings reveal significant misclassification rates across all indigenous soil fertility assessment approaches investigated, with misclassification levels of 71%, 67%, 66%, and 61% for vegetation cover, weeds, soil colour, and previous yield, respectively. This suggests that while indigenous methods such as weeds, previous yields, vegetation cover, and soil colour are widely used, they may lack the precision needed to accurately measure soil fertility, particularly when compared to scientific testing. Indigenous approaches are often based on farmer experiences and observations, but their accuracy in soil fertility assessment is limited. This is consistent with findings from other studies that have highlighted the inefficiency of the indigenous approaches, especially in predicting soil fertility based on soil pH, nitrogen levels, phosphorus, and potassium [16, 34, 46].

Amongst all the indigenous approaches investigated, the previous yield indicator presented a slightly reliable soil fertility classification, though it was also not immune to error. Previous yields might be influenced by factors beyond soil fertility, such as weather conditions, pest infestations, or inputs like fertilizers and pesticides [1]. However, in cases where smallholder farmers are resource-constrained, the study suggests that complementing previous yield assessments with periodic soil testing could offer a realistic approach to improving soil fertility management among smallholder farmers. This approach balances the cost of frequent soil testing with the affordability of relying on past yield indicator soil fertility. Moreover, Buthelezi et al. [7] reiterated the need to integrate indigenous knowledge with scientific tools to provide more holistic and adaptive management practices for resource-poor farmers.

Nevertheless, from the observed misclassification levels across the indigenous soil fertility approaches, reinforcing soil testing emerges as a crucial intervention to improve soil fertility management. Recent advancements in soil testing technologies in Uganda, such as AgroCares and Soil Cares, have introduced portable, real-time soil analysis tools that provide immediate feedback on soil fertility levels to farmers. These technologies offer affordable and user-friendly solutions that can be integrated with the previous yield indicator. With these innovations, smallholder farmers can significantly enhance their ability to make informed decisions about land management, especially fertilizer application [33].

After examining the level of soil fertility misclassification, we assessed the correlates of misclassification for each indigenous approach. Notably, the plot size and land tenure (being a land owner) were found to be positively associated with soil fertility misclassification. Larger plot sizes were associated with a higher likelihood of misclassification, which could be attributable to the variability in soil characteristics within larger plots. Farmers might struggle to apply consistent observational techniques across diverse sections of their land, leading to overgeneralization or misinterpretation of soil fertility indicators. This is similar to findings by Tittonell et al. [36], where spatial heterogeneity in soil fertility within larger plots often led to inaccurate assessments by farmers in Kenya and Uganda. Furthermore, in an assessment of determinants of farmer’s knowledge of soil and water in Kenya, Njenga et al. [27] also found that the smaller the plot size, the more knowledgeable the farmers were. This implies that for larger plots, it is critical to invest in scientific soil testing to gain a more accurate understanding of soil fertility since all the indigenous approaches are ineffective.

Land ownership also positively influenced misclassification of soil fertility, suggesting that farmers who own land have different perceptions about their soil fertility compared to those who rent it. This could result from landowners who have lived on or worked on a piece of land for a long time having a strong attachment to the land or having overconfidence in their Indigenous soil assessment approaches leading to errors. Farmers who rent land also tend to be more critical and evaluate the soil fertility of the land in question whereas the land owners tend to market their land [18]. However, this finding contradicts that of Sklenicka et al. [35], who highlighted that landowners are often more knowledgeable about the fertility of their soil compared to those who rent land.

5 Conclusions

Whereas indigenous soil fertility assessment approaches are cost efficient for smallholder farmers especially in Sub-Saharan Africa, they are prone to errors. This underscores the urgent need for promotion of affordable and portable soil testing kits. The previous yield indicator performed better than the vegetation cover, soil colour, and weeds indicators. However, it also has a high level of inaccuracy due to high variation of soil fertility within the plots, hence necessitating periodic scientific soil testing.

This study also assessed the correlates of soil fertility misclassification using a binary probit model and a multivariate probit model. The binary probit results showed the plot size significantly increased the likelihood of soil fertility misclassification across all the indigenous approaches assessed. Being a land owner, only increased the likelihood of misclassification based on the vegetation cover. The multivariate probit results showed that the age of the farmer, access to extension services, plot size, and being a land owner significantly influenced the likelihood of misclassification.