The urban commune of Gaya is located in the extreme south of Niger, within the Dosso region, bordering both Benin and Nigeria. Its strategic transboundary position makes it a key agricultural and commercial hub in the country’s southern corridor. Gaya lies within the Sudanian Sahelian agroecological zone, characterized by a prolonged dry season and an average annual rainfall ranging from 700 to 900 mm [15]. The presence of the Niger River provides relatively stable access to water resources, enabling the development of irrigated rice farming, particularly during the dry season when climatic risks are lower and irrigation control is higher [16]. The commune hosts several irrigated perimeters such as Karé Gourou, Tanda, and Bengou where rice cultivation is practiced intensively. These perimeters are equipped with semi-modern hydro-agricultural infrastructure, including pumping stations, drainage systems, and water distribution networks, although maintenance levels vary across farmer groups. Rice farming in Gaya mobilizes significant labor and contributes substantially to rural household incomes, positioning the commune as a focal point for seasonal agricultural activity. Socioeconomically, Gaya is dominated by family-based farming systems, often organized into cooperative groups. The diversity of farming practices, ranging from traditional to semi-intensive systems, reflects varying levels of access to inputs, technical knowledge, and institutional support. This heterogeneity makes Gaya an ideal site for analyzing technical efficiency and modeling agricultural performance.

This study draws on primary data collected from 103 rice farms selected from a population of 731 producers operating within the major irrigated perimeters of the urban commune of Gaya (Karé Gourou, Tanda, and Bengou). The sample size was determined using a probabilistic sampling formula (1) to ensure adequate representativeness of the target population. A stratified random sampling approach was then applied: the total population of 731 producers was first stratified by irrigated perimeter (Karé Gourou, Tanda, and Bengou) to reflect structural heterogeneity. Within each stratum, farms were randomly selected proportionally to the size of the farmer population in that perimeter, until the target sample of 103 farms was reached. This procedure ensured that each farm had an equal probability of selection while maintaining representativeness across the three irrigated Data were collected during the dry season through a structured questionnaire administered digitally using Kobo Collect platform, and subsequently processed and analyzed in R. The survey captured detailed information on the socioeconomic characteristics of farm managers, the mobilization of production factors, and the economic performance of each farming unit. Data collection was conducted in two stages. The first stage involved focus group discussions with representatives of local farmer organizations. The focus group discussions were analyzed using a thematic content analysis approach. Notes and transcripts from the sessions were coded inductively to identify recurring themes related to infrastructure, governance, gender composition, and nursery management practices. These themes were then organized into categories (access to inputs, cooperative governance, irrigation challenges) and triangulated with survey data to provide contextual depth. This qualitative analysis served two purposes: 1) to refine the questionnaire by incorporating locally relevant variables, and 2) to interpret quantitative efficiency results within the broader institutional and organizational context of irrigated rice farming in Gaya. The second stage consisted of individual interviews with rice farm managers, who constitute the primary unit of analysis. The margin of error was set at 9% to balance statistical precision with practical feasibility. While a 5% margin would have required a substantially larger sample size, the 9% threshold allowed us to obtain a representative sample of 103 farms across the three irrigated perimeters, within the logistical and resource constraints of fieldwork. This level of precision is consistent with accepted practice in exploratory agricultural studies where margins between 5%-10% are considered adequate.

Target population (N): 731 (total rice farmers,provided by the farmers organization)Confidence level (tp): 1.96 (corresponding to 95%)Estimated proportion (P): 0.5 (to maximize variance)Margin of error (y): 9% (0.09)

2.3.1. Typology of Rice Farming Systems

A farm typology was developed to reduce structural heterogeneity and ensure more meaningful comparisons among production units. The typology was constructed through a Principal Component Analysis (PCA) applied to key structural, technical, and socioeconomic variables (Table 1), followed by a Hierarchical Ascendant Classification (HAC) using Ward’s method to derive homogeneous clusters of farms. The selected variables capture essential features of production systems, including input intensity, land and labor resources, and household characteristics. Socioeconomic attributes such as education level, gender of the farm manager, and income diversification were incorporated because they reflect managerial capacity and household resilience factors known to indirectly influence production performance. For example, education may enhance input allocation decisions, while income diversification can reduce exposure to production risks and stabilize resource use. The combined PCA–HAC approach provides a robust multivariate framework for identifying structurally consistent groups of farms.

Table 1.Variables Used in the combined PCA-HAC approach

2.3.2. Hybrid DEA-Machine Learning Approach

The hybrid DEA-Machine Learning (DEA-ML) approach combines the complementary strengths of production econometrics. While Data Envelopment Analysis (DEA) provides a robust and economically interpretable measure of technical efficiency (2), Machine Learning algorithms enable the identification, ranking, and prediction of complex determinants of efficiency in nonlinear and multidimensional contexts.

The input variables in this DEA model include:

Cultivated land area (hectares),Quantity of seeds (kg),Quantity of fertilizer (kg),Quantity of urea and herbicides (kg),Irrigation fees,Miscellaneous expenses,Labor required during the different phases of rice cultivation (pre-rice, rice, and post-rice).

The final output is measured by the total rice production (kg). DEA relies on two main models: the CCR model and the BCC model (Charnes et al. 1978). The CCR model assumes constant returns to scale and the BCC model allows for variable returns to scale, which is often more suitable in agriculture where productivity may vary with farm size [17].

2.3.3. Integration of Socio-Economic Factors

Socio-economic variables such as the farm manager’s education level, marital status, and the number of active household members can also be incorporated into the DEA analysis to better understand their influence on farm efficiency. Previous studies have shown that higher levels of education and more structured family labor organization are associated with better resource utilization and more efficient farm management [18].

Machine Learning into the DEA framework enhances both the explanatory and predictive power of efficiency models. According to the support of [19], the hybrid approach can be implemented through two complementary strategies: 1) ML-Enhanced DEA: Machine Learning algorithms are applied upstream to select relevant explanatory variables, reduce dimensionality, or correct noisy data prior to the implementation of DEA. This preprocessing step improves the robustness and interpretability of efficiency estimates by refining the input space and minimizing distortions caused by outliers or irrelevant features. 2) DEA-Informed ML: DEA efficiency scores are used as target variables in supervised learning models to identify nonlinear and interaction-based determinants of performance. This integrated approach enables the capture of the multidimensional complexity of agricultural systems, which are characterized by strong interdependencies among biophysical, socioeconomic, and environmental inputs. Furthermore, ML models offer the possibility of conducting sensitivity analyses and feature importance assessments, thereby highlighting the most influential factors driving farm-level efficiency. The analytical process is structured into three successive stages:

Step 1: Farm Segmentation

In the first stage, farm heterogeneity is explicitly addressed through an unsupervised classification approach. Hierarchical Ascendant Classification (HAC), following dimensionality reduction via Principal Component Analysis (PCA), is applied to a set of structural, agronomic, and socio-economic variables in order to identify homogeneous groups of rice farms. This segmentation reduces intra-group variability and improves the comparability of decision-making units within each group, thereby enhancing the precision and robustness of subsequent efficiency estimates. Such a typology-based approach is particularly recommended in heterogeneous production environments, where pooling structurally different farms may bias efficiency measurement [20].

Step 2:Intra-Group Efficiency Estimation

Given the context of Gaya, which is marked by limited access to productive resources characterizing smallholder farming systems, an input-oriented DEA model was adopted. This approach is consistent with empirical studies conducted in West African agricultural contexts, where producers primarily seek to optimize the use of scarce inputs in order to sustain or increase output levels [21]. For each homogeneous group, DEA scores (θᵢ) are estimated separately using the input-oriented Variable Returns to Scale (VRS) model. This multi-level approach ensures that performance measurement is tailored to the internal structure of each farm type [18]. In addition to the input‑oriented DEA specification, which is appropriate in resource‑constrained contexts where farmers seek to minimize (3) input use, an output‑oriented DEA model can also be applied for robustness. The output‑oriented approach evaluates the potential proportional increase in rice production that each farm could achieve with its current level of inputs. Comparing results across orientations allows us to verify whether efficiency scores remain consistent and to identify whether constraints are primarily input‑related or output‑related. This dual specification strengthens the robustness of the analysis and provides a more comprehensive understanding of farm performance.

subject to:

où:

Y_i represents the output vector (rice production),X_i the input vector (land area, seeds, fertilizer, labor, herbicide, and urea),λ_j the weights assigned to the reference farms,and θ ∈[0,1], the technical efficiency score of farm i.

A value of θ = 1 indicates a technically efficient farm (located on the production frontier), whereas a value of θ < 1 reflects a potential for improvement equivalent to 1 − θ expressed as the proportion of input reduction required to reach the frontier.

Step 3: Predictive Modeling of Efficiency Determinants

The technical efficiency scores obtained (θ_i) (4)are subsequently used as the target variable in a supervised Machine Learning model, specifically a Random Forest algorithm. This model aims to identify and rank the key determinants of efficiency across heterogeneous farming systems. The explanatory variables include technical inputs, as well as the socioeconomic and demographic characteristics of farm managers. The general form of the predictive model is expressed as:

where:

θ_i: Technical efficiency score of farm i, X_i: Technical inputs for farm i,Z_i: Socio-demographic and economic characteristics of farm manager,ϵ_i: Error term, representing unmodeled factors that influence efficiency,f(X_i,Z_i): The predictive function is a complex mapping of the explanatory variables X_i and Z_i for each farm ii, where f(⋅) denotes the function learned by the algorithm. In this context, X_i represents the technical inputs, and Z_i captures the socio-demographic and economic characteristics of the farm manager. The Random Forest algorithm models this relationship to estimate the technical efficiency score θ_i.

This modeling approach allows for the exploration of nonlinear relationships and interaction effects that may not be captured by traditional econometric methods. Moreover, the Random Forest algorithm provides variable importance metrics, enabling the identification of the most influential factors driving efficiency at the farm level. Variable importance metrics derived from the model such as the increase in mean squared error (MSE) or node purity enable the identification of key drivers of efficiency. These indicators highlight the most influential factors, providing actionable insights into the levers that can improve farm performance.

2.3.4. Random Forest Model

The Random Forest algorithm operates by constructing an ensemble of independent decision trees to predict the technical efficiency score θ_i (5). For each tree, a random subset of explanatory variables is selected, and predictions are made through successive decision splits. The final output of the Random Forest model is the average of the predictions generated by all individual trees, thereby reducing variance and improving generalization performance.

where:

θ ^ i : Prediction of the technical efficiency score for farm i,N: Number of trees in the forest, f j ( X i , Z i ) : Prediction of tree j for farm i.

The Random Forest model typically minimizes the Mean Squared Error (MSE) to adjust the weights of the trees within the forest:

where:

n is the total number of farms in the sample,θ_i is the actual efficiency score for the farm i, θ ^ i is the predicted efficiency score for farm i.

The model predicts each farm’s capacity to reduce input use while maintaining its current level of rice production. This reflects the potential for technical efficiency gains without compromising output. The explanatory variables enabled the identification of key factors influencing the technical efficiency of rice farms, offering insights into the structural and managerial levers that can enhance performance.

Contributions of theVariables

To account for structural heterogeneity among rice farms in Gaya, a Principal Component Analysis (PCA) was conducted using 17 agronomic, economic, and socio-demographic variables. The first three principal components explained 54.16% of the total variance, with eigenvalues of 5.83, 2.59, and 1.33 respectively (Table 2). These components revealed distinct dimensions of farm organization and resource use. The first dimension, which accounted for 32.41% of the total variance, was primarily shaped by agronomic and economic variables. High contributions were observed for fertilizer quantity (14.40%), production output (13.60%), land area (13.58%), total input cost (12.29%), seed quantity (11.45%), irrigation fees (10.19%), and urea application (9.92%). This dimension reflects a gradient of input intensity and productive capacity, distinguishing farms with higher resource allocation and output levels. The second dimension, explaining 14.38% of the variance, was dominated by labor-related variables. The most influential were labor during the rice phase (28.19%), post-rice phase (23.73%), pre-rice phase (21.11%), and the number of active household members (13.30%). This axis captures the structure and timing of labor deployment across production stages. While labor intensity varied significantly across farms, its contribution to technical efficiency was modest, suggesting that labor quantity alone is insufficient without strategic allocation and timing. The last Dimension, accounting for 7.37% of the variance, was defined by socio-demographic and residual economic traits. Education level (25.37%), age of the farm manager (14.34%), number of income sources (13.99%), gender of the farm manager (10.19%), miscellaneous expenses (11.37%), and marital status (7.59%) were the leading contributors. This component reflects household diversity and financial behavior.

Table 2. Variable contributions to the first three principal components.

Hierarchical Ascendant Classification

The Hierarchical Ascendant Classification (HAC), implemented using Ward’s method, identified three homogeneous clusters of rice farms based on their structural and socioeconomic characteristics:

Cluster 1—Commercial or Semi-Intensive Farms (F1): This group comprises relatively large farms managed by experienced producers. These farms exhibit high levels of input use, more intensive production practices, and stronger integration into commercial markets.Cluster 2—Small Family Farms (F2): These farms are small in size, rely predominantly on family labor, and use inputs more sparingly. Production practices are generally traditional, and resource constraints are more pronounced than in other groups.Cluster 3—Medium-Sized Farms (F3): Representing the largest share of the sample, this cluster includes farms of intermediate size, with an average cultivated area of 0.31 hectares. Their input use and production intensity fall between the two previous groups.

These clusters reflect distinct production logics and resource endowments within the irrigated perimeters of Gaya. They provide a meaningful basis for analyzing efficiency differentials and for designing targeted interventions aimed at improving farm performance (see Figure 1).

Table 3 indicates that the majority of farm managers lack formal education across all farm categories (67.2% in medium‑sized farms, 64.3% in commercial farms, and 68.4% in small family farms), with no statistically significant differences between groups (χ² = 4.540; p = 0.338). This pervasive absence of formal schooling reflects structural constraints in rural Niger, where agricultural knowledge is predominantly transmitted informally through family and community networks rather than vocational or technical training. Similarly, marital status and gender distributions do not vary significantly across farm types (p > 0.05), with most managers being married men, underscoring the persistence of male dominance in land access and decision‑making within irrigated perimeters. In contrast, household income diversification differs significantly across farm categories (χ² = 16.42; p = 0.012). Medium‑sized farms demonstrate greater diversification, combining rice cultivation with complementary activities such as petty trade, livestock rearing, and seasonal migration. These findings suggest that income diversification may represent a critical pathway for enhancing economic sustainability in rice‑based systems.

Figure 1. Rice Farms groups.

Table 3.Socioeconomic characteristics of farm managers by rice farm type

Figure 2 illustrates that commercial or semi-intensive farms (Cluster 1), characterized by larger landholdings, make substantial use of inputs such as seeds, fertilizers, and urea. This intensive input use enables them to achieve high rice yields ranging from 2000 to 6000 kg, with some farms reaching up to 9000 kg. These farms are typically managed by older farm heads, reflecting greater agricultural experience. In contrast, small family farms (Cluster 2), with smaller plots, rely on moderate input levels and produce lower rice yields, generally between 1500 and 4000 kg. These farms are often managed by younger producers and follow less intensive, more traditional farming practices. Medium-sized farms (Cluster 3) represent a middle ground, with moderate land area and input use, yielding between 3000 and 6000 kg of rice. The farm managers in this group tend to be of intermediate age, suggesting a balance between traditional knowledge and innovative management approaches.

Figure 2.Technical indicators by rice farm typology.

Figure 3 shows that commercial or intensive farms (Cluster 1) exhibit the highest technical efficiency scores (average DEA score of 0.83) and the largest share of efficient units (27%). Their median score surpasses that of medium‑sized farms (0.76) and small family farms, reflecting the advantage of larger landholdings, higher input intensity, and greater managerial experience. In the context of Gaya, these farms benefit from better access to irrigation infrastructure and markets, which allows them to optimize resource use more effectively. In contrast, small family farms (Cluster 2) display greater variability and relatively lower DEA scores (average of 0.71), with only 12% of units classified as efficient. This lower performance may be linked to constraints in input access, reliance on family labor, and limited financial capacity. Such conditions often lead to suboptimal resource allocation and reduced resilience to shocks, highlighting the structural vulnerability of smallholders in Gaya’s irrigated perimeters. Medium‑sized farms (Cluster 3) occupy an intermediate position, with balanced input use and production levels, and a moderate efficiency profile. Their average score and share of efficient units suggest a more stable and adaptive management approach. In Gaya’s transboundary economy, these farms often combine traditional practices with incremental innovations, enabling them to maintain efficiency despite resource constraints. This intermediate performance underscores their potential as a target group for policy interventions, as they are well positioned to scale up efficiency gains if provided with improved access to inputs, training, and institutional support.

Figure 3.Distribution of DEA scores by farm type and share of efficient units.

The results reveal a strong concordance between observed and predicted DEA scores, as evidenced by a high coefficient of determination (R² = 0.9168) and a very low Mean Squared Error (MSE = 0.0027). To validate model robustness, we report the Out‑of‑Bag (OOB) error estimate, which provides an internal cross‑validation during Random Forest training. The OOB error rate was 0.087, corresponding to a mean squared error (MSE) of 0.0031, confirming that predictive accuracy remained stable when tested on unseen data. As an additional check, a 5‑fold cross‑validation was performed, where the dataset was partitioned into five subsets and iteratively trained and tested. The average cross‑validated R² was 0.89, with a mean absolute error (MAE) of 0.041, closely aligned with the original model performance. These results demonstrate that the Random Forest model generalizes well and is not overfitting the sample, thereby strengthening confidence in the identified determinants of technical efficiency. The correlation between observed and predicted scores reaches 0.9575, indicating the model’s exceptional ability to capture efficiency variations based on the selected explanatory variables. The dispersion of points around the ideal diagonal line (y = x) remains minimal, further confirming the model’s precision and stability in predicting technical efficiency across rice farms (Figure 4). From an economic perspective, this high level of predictive accuracy suggests that the Random Forest model successfully identifies the structural and managerial factors that drive efficiency in Gaya’s irrigated rice systems.

Figure 4.Observed vs. predicted DEA scores using the random forest model.

The Random Forest regression model was employed to complement the DEA results by identifying the determinants of technical efficiency among family rice farms. Variable importance was assessed using the increase in mean squared error (IncMSE) and node purity indices, which quantify the predictive contribution of each factor to efficiency scores. The analysis revealed that seed quantity (17.44) is the most influential variable, confirming that the choice and dosage of seeds are central to production efficiency. This finding aligns with agronomic literature emphasizing the role of seed quality and planting density in yield optimization. Farm size (7.63) ranked second, suggesting that scale effects and land allocation strategies significantly influence efficiency, consistent with the hypothesis of economies of scale in smallholder systems. Other agronomic inputs also demonstrated positive contributions. Herbicide (6.25) use and fertilizer (6.08) application were strong predictors, highlighting the importance of weed control and nutrient management. The ability to optimize direct costs and production-related fees (5.57) contributed positively, indicating that farms which strategically allocate financial resources achieve higher efficiency. Urea application (3.14) further reinforced the role of targeted fertilizer use. Additionally, household income diversification (1.18) and family labor during the rice production phase (0.11) showed modest but positive effects, suggesting that economic resilience and timely labor allocation can support technical performance. In contrast, socio-demographic variables such as marital status, education level, and gender of the household head exhibited negligible or negative importance. Similarly, labor-related variables outside the production phase contributed little to efficiency, implying that efficiency is more dependent on the quality and timing of input use than on household composition.

Table 4. Variables Importance and theirs ranks.