A DEA-based evaluation of upstream food supply chain efficiency: evidence from Australian banana farms

The global food supply chain (FSC) faces increasing pressure to enhance sustainability, reduce waste and improve food security (FAO, 2023; Gardas et al., 2019; Gurrala and Hariga, 2022). This urgency is driven by the alarming reality that more than one-third of global food production is lost or wasted annually (FAO, 2023; Pfaltzgraff et al., 2013), much of it within the FSC itself (Chauhan, 2020; Gustavson et al., 2011; UN, 2024). The consequences are profound: over 783 m people endure chronic hunger each year and around 150 million children suffer from stunted growth due to food scarcity and nutrient deficiencies (FAO, 2023; UN, 2024).

Numerous studies have explored ways to reduce food loss and waste (Luo et al., 2023; Magalhães, 2021). Among the most waste-prone products is the banana, an essential fruit with high nutritional value and significant economic relevance (Chauhan, 2020; FAO, 2023; Giuggioli et al., 2024). Despite being one of the world’s top-traded fruits, bananas are highly perishable and particularly vulnerable to supply chain inefficiencies, especially due to their sensitivity to temperature and handling (Chauhan, 2020; Slatter, 2023).

In Australia alone, banana production reached 374,251 tonnes in 2023, worth $583m (Hort Innovation, 2024). Yet, losses remain substantial. The 2022–23 ABARES Horticulture Survey reported an average crop loss and waste of 24% per banana farm (Slatter, 2023). If just one-third of this waste could be avoided and converted into marketable produce, an additional 39,395 tonnes, valued at $61.4m, could enter the supply chain. These figures highlight the scale of preventable waste and underscore the urgent need for more efficient and sustainable practices (Fabi et al., 2021; Yuvaraj et al., 2023).

In the context of the FSC, waste is often the result of inefficiencies across production, storage, distribution or consumption stages (Parfitt et al., 2010; Papargyropoulou et al., 2014). It reflects resource misuse, including labour, land, energy, water and logistics, resulting in lower useful output. Thus, food waste can be seen as a direct proxy for inefficiency. Treating waste as a negative performance outcome, it is crucial to apply robust analytical methods that benchmark performance, identify inefficiency sources and guide targeted improvements. This is especially important in agricultural systems, where production environments are heterogeneous and outputs are highly variable.

This study focuses on banana waste generated during farming, processing and packaging stages as an indicator of supply chain inefficiency. Efficiency at these early stages is particularly critical, as they account for a significant portion of total food loss (FAO, 2023; Giuggioli et al., 2024; Harvey et al., 2016). Improvements in input utilisation and farm management can reduce production costs, increase yields and enhance both environmental outcomes and grower profitability.

In Australia, banana crop losses have been linked to inefficient use of inputs and overproduction driven by certain agronomic practices and large-scale production systems (Goodwin, 2023; Messner et al., 2021). Losses were estimated at 29% in 2021–2022 and 24% in 2022–2023 (Downham, 2022; Slatter, 2023), with consistent findings from Central Queensland University suggesting a 20–29% range across years (CQU, 2024). While growers cannot always control market demand or output prices, they can take steps to reduce waste and improve yields by optimising fertiliser and chemical use, water management, packaging and post-harvest processes (Harvey et al., 2016). Top-performing farms demonstrate that greater efficiency leads to higher net returns and better resource utilisation.

In addition to government and institutional support through training programmes, demonstration farms and expert advice (Agriculture Victoria, 2024), peer-to-peer learning also offers promise. Growers' cooperatives and farmer groups can facilitate shared knowledge, joint training and practical learning exchanges (Grower Group Alliance, 2023; Hort Innovation, 2020). However, widespread benchmarking remains limited. While some industry reports exist (e.g. Hort Innovation, 2015), few tools are available that allow individual growers to compare their efficiency to peers in a practical, robust and accessible way. Growers are also often reluctant to share operational details due to competitiveness and confidentiality concerns. This makes the development of simple but effective benchmarking tools essential. Such tools could help grower associations assess individual farm efficiency and identify high performers as peer benchmarks. Efficiency measurement, in this context, becomes a vital element of performance management in FSCs.

Efficiency is commonly assessed using either parametric or non-parametric methods. Parametric approaches, such as stochastic frontier analysis, statistically estimate a production frontier and separate inefficiency from random noise (Kumbhakar and Lovell, 2000). While these models allow for hypothesis testing, they require predefined functional forms and strong assumptions about data distribution, which are often unsuitable for small-scale or diverse farming environments.

Non-parametric methods, such as data envelopment analysis (DEA) and the Malmquist Productivity Index (MPI), offer an alternative. DEA constructs an efficiency frontier using linear programming and measures a farm’s performance relative to this benchmark (Charnes et al., 1978, 1981). It handles multiple inputs and outputs without needing predefined functional forms and is well suited to cross-sectional analyses of performance. MPI, on the other hand, measures productivity changes over time but requires reliable time-series data, which may not always be available or interpretable by non-specialists.

Given the cross-sectional nature of the available data and the need to benchmark across farms at a single point in time, this study applies DEA to assess banana farm efficiency. DEA defines a farm’s efficiency as the ratio of the weighted sum of its outputs to the weighted sum of its inputs (Thanassoulis et al., 2008). Farms operating on the efficiency frontier receive a score of 1 (100%) and serve as benchmarks for those with lower scores.

However, traditional DEA has a known limitation: it assigns weights that maximise each farm’s efficiency score, which can result in some inputs or outputs receiving zero weight and being effectively ignored (Cooper et al., 2011). To overcome this, the study also employs entropy-based DEA, which integrates entropy weighting to assign more objective and data-driven weights (Xie et al., 2014). The entropy method uses the variability of each variable to determine its informational value, preventing the exclusion of potentially important factors (Roszkowska et al., 2024). By incorporating Shannon's (1948) entropy concept, the model better reflects data diversity and enhances robustness.

Entropy-based DEA thus addresses the shortcomings of traditional DEA by improving discriminatory power, ensuring all variables are considered, and accommodating the complexity of horticultural data. This dual approach enhances the reliability of the findings and provides a strong foundation for practical benchmarking, making it more suitable for complex, multi-variable performance evaluations. In this study, we raise the following research question:

To answer it, we apply both traditional and entropy-based DEA to measure the farm-level efficiency of Australian banana growers. The goal is to identify whether farms are operating efficiently in resource utilisation and waste reduction or whether inefficiencies exist that can be addressed through improved management practices. By benchmarking the most and least efficient farms, growers can learn from best-practice peers, while policymakers and industry leaders gain insights for shaping sustainability strategies and support mechanisms.

The paper is organised as follows: Section 2 reviews related literature on efficiency measurement in FSCs. Section 3 outlines the methodology and data used. Section 4 presents the results of the DEA analysis. Section 5 discusses the strategic implications for improving the sustainability of banana farming. Section 6 concludes with limitations and directions for future research.

Assessing agricultural production efficiency is fundamental to developing strategies that reduce food loss and waste, alleviate rural poverty and enhance farming sustainability (Ait Sidhoum et al., 2020; Atici and Podinovski, 2015; Paul, 2024; Singh and Chhetri, 2019). In Australia, analyses by Food Innovation Australia Limited (FIAL) reveal persistent inefficiencies and substantial food waste across agri-food supply chains, particularly in horticulture. FIAL’s National Food Waste Strategy Feasibility Study identifies significant banana losses at production, distribution and retail stages, largely driven by uneven quality standards, fragmented supply chains and logistics constraints (FIAL, 2021).

This study builds on these findings by conceptualising food waste as a direct indicator of inefficiency within a DEA framework, aligning with FIAL's national priorities on waste reduction and sector competitiveness. Each farm or supply chain node is treated as a decision-making unit (DMU) that converts inputs into outputs under resource and operational constraints, reflecting the foundational supply chain logic that node-level efficiency shapes overall network performance (Charnes et al., 1978; Cooper et al., 2007; FIAL, 2021).

While traditional DEA models consider only desirable outputs, environmental extensions integrate undesirable outputs such as emissions or food waste to capture inefficiencies more comprehensively (Seiford and Zhu, 2002; Zhou et al., 2008). Here, food waste is modelled as a non-desirable output, representing inefficiencies from resource mismanagement (land, water, fertiliser, labour and energy) and operational failures such as poor coordination, inadequate storage and inaccurate demand forecasting (Gruber et al., 2016; FIAL, 2021). This approach aligns with reverse logistics and circular economy principles that seek to minimise waste while maximising productive outcomes (Seiford and Zhu, 2002; Zhou et al., 2008).

The modelling further applies the Directional Distance Function (DDF) and weak disposability principles, enabling simultaneous improvement in desirable outputs and reduction in undesirable ones (Chung et al., 1997). This framing reflects contemporary research recognising that inefficiencies at individual nodes can propagate across entire supply chains.

Globally, productivity and food waste reduction have been enduring policy concerns since the founding of the Food and Agriculture Organization of the United Nations (FAO) in 1945 (Fabi and English, 2019; Parfitt et al., 2010). In Australia, these priorities remain central, as evidenced in FIAL’s benchmarking and sector insights reports (FIAL, 2021). Recent global initiatives increasingly emphasise efficiency optimisation across all stages of the agri-food supply chain to mitigate food insecurity linked to systemic loss, waste and uneven distribution (Fabi et al., 2021; Yuvaraj et al., 2023; FIAL, 2021). Table 1 summarises major stages in the banana supply chain where waste occurs, along with key causes and impacts.

Identifying and addressing inefficiencies allows for more targeted resource allocation and strengthens the sustainability of supply chain operations (Paul, 2024; Şahin et al., 2021; Singh and Gundimeda, 2021). Beyond merely quantifying losses, efficiency analysis reveals underlying weaknesses in production systems and informs specific improvement strategies, making it a critical tool for advancing sustainable operational performance (Kyrgiakos, 2023).

In the agricultural sector, DEA has been widely used to assess farm efficiency and enhance productivity and competitiveness (Ho et al., 2022; Košařová and Pokrivčák, 2023; Nandy and Singh, 2020; Paul, 2024; Wagan et al., 2018). Its widespread adoption reflects the critical role of efficiency in supporting strategic farming activities, e.g. fostering collaboration between smaller, less efficient farms and larger ones to optimise input use and improve outcomes.  Appendix 1 provides additional details on the application of DEA for efficiency evaluation in the agricultural sector and outlines the commonly used DEA approaches.

DEA also facilitates benchmarking, allowing farms to identify performance gaps and adopt best practices to reduce waste and enhance sustainability (Košařová and Pokrivčák, 2023; Paul, 2024). Benchmarking is particularly useful in contexts where the DMUs are relatively homogenous, as is the case in the Australian banana industry. Large banana farms are concentrated in North Queensland, benefiting from economies of scale and infrastructure access, while smaller and medium-sized farms, located in Northern New South Wales and Western Australia, cater to local and niche markets (ABGC, 2025b).

Although systemic issues, such as overproduction and waste driven by contractual obligations and strict cosmetic standards imposed by powerful retailers, require long-term structural reforms, more immediate improvements are achievable. These include enhancing farm management skills and input use efficiency to reduce food loss and waste (Goodwin, 2023).

Agricultural efficiency studies commonly use input-oriented DEA models, as farmers typically have more control over inputs than outputs, which are subject to weather, pests and market fluctuations. Inputs often include costs or quantities of labour, irrigation, fertiliser, packaging, chemicals and machinery maintenance (Lin et al., 2019; Kumar and Kumar, 2016). In banana farming, efficient allocation of such inputs directly affects yields (e.g. cartons packed per hectare) and waste outcomes.

Despite DEA’s potential, its application to banana farming remains limited, largely due to the difficulty of collecting primary farm-level data. Most studies rely on secondary sources like industry yearbooks (Lin et al., 2019; Molina et al., 2023). While many focus on comparing performance across farms (e.g. Ho et al., 2022; Paul, 2024), fewer studies explore inefficiencies linked to management or sustainability. Notable exceptions include Picazo-Tadeo et al. (2011) and Ait Sidhoum et al. (2020), who identified significant inefficiencies and sustainability gaps in European farming systems.

This study applies both traditional and entropy-based DEA to evaluate Australian banana farm efficiency. The entropy-based approach enhances DEA’s discriminatory power by assigning objective weights to inputs and outputs, thus reducing the number of farms incorrectly deemed fully efficient (Xie et al., 2014). Tables 2 and 3 compare the methods and summarise their agricultural applications.

Although DEA primarily quantifies relative efficiency by comparing input-output performance across DMUs, the insights derived from efficiency benchmarking can directly inform strategies to reduce crop loss and food waste. In agricultural production systems, inefficiency often reflects suboptimal resource allocation, operational mismanagement or technological gaps – factors that also contribute to post-harvest losses and waste (Xue et al., 2017; FAO, 2019). By identifying input redundancies and underperforming practices, DEA results can guide farmers in optimising the use of land, fertiliser, labour and energy, thereby improving yield quality and reducing avoidable loss during production and handling (Yang et al., 2023; Gadanakis et al., 2015).

Improved resource efficiency not only minimises unnecessary input use but also enhances process stability and output consistency, both of which are critical for reducing on-farm grading losses and post-harvest deterioration (Parfitt et al., 2010; FIAL, 2021). When interpreted through a supply chain management lens, DEA outcomes can also facilitate peer learning and benchmarking among growers. These can include encouraging the adoption of best practices in irrigation scheduling, fertiliser use and harvest timing that mitigate overproduction and spoilage (Reutter et al., 2017; Babu and Blom, 2014). Thus, while DEA efficiency scores are quantitative, their qualitative implications extend to actionable pathways for reducing food waste by improving operational discipline and resource stewardship across the supply chain.

In traditional DEA, two core models are used to estimate efficiency: the Constant Returns to Scale (CRS) model and the Variable Returns to Scale (VRS) model. Each model constructs its own efficiency frontier. Firms or decision-making units (DMUs), that lie on the frontier are deemed 100% efficient, while those beneath it are inefficient to varying degrees.

The CRS model assumes all firms operate under constant returns to scale, i.e. output changes proportionally with input. This model presumes a perfectly competitive environment where firms can freely scale up or down without affecting productivity, which is rarely the case in practice. In contrast, the VRS model acknowledges that firms often face constraints preventing them from operating at an optimal scale. It reflects real-world production conditions that may exhibit increasing or decreasing returns to scale, thereby providing a more flexible and realistic evaluation of efficiency.

Since there are farms of different scales in the Australian banana-growing business and it is easier for the growers to control the input in terms of fertiliser and chemical usage, water management and packaging, a VRS DEA model with input orientation is employed in this research. The mathematical formulation is given in  Appendix 2.

Traditional DEA models assume equal importance (or weight) for all input and output variables and evaluate the relative efficiency of DMUs by optimising weights. However, this can lead to biased or unrealistic results when all variables are treated with equal weight, regardless of their significance or variability. Shannon's (1948) entropy is a measure of uncertainty or variability in data. In the context of DEA, the entropy-based weighting mechanism assigns weights based on the variability or discriminative power of each variable. Variables with higher variability and more informative content are given higher weights, as they provide more meaningful distinctions between DMUs. This ensures that variables with significant contributions to efficiency are emphasised over less impactful variables. Entropy-based DEA enhances the discriminatory power of the analysis by reducing instances of multiple DMUs being evaluated as fully efficient (i.e. efficiency scores of 1). This is particularly important in applications with numerous DMUs, as it allows for a finer differentiation of performance levels (Xie et al., 2014).

Entropy-based DEA models are the same as those of traditional DEA except the weights of the input and output variables are calculated based on Shannon's (1948) entropy-based weighting mechanism. The steps are shown in  Appendix 3.

In Australia, bananas are grown in both tropical and subtropical regions. As such, the industry is diverse in terms of the geographical location of banana farms, farming practices, the size and type of farms that grow bananas, the varieties of bananas grown and their flavour (Akbar et al., 2024). Figure 1 shows a schematic map of banana-producing regions in Australia.

This study analyses the efficiency of 27 Australian banana farms located across Queensland, New South Wales and Western Australia. Due to intense market competition, growers are generally reluctant to share sensitive operational data. Despite efforts to collect primary data through the Australian Banana Growers’ Council and at the 2024 Australian Banana Industry Congress, participation was limited. As a result, this research draws on secondary data from the Banana Enterprise Performance Comparison report published by Horticulture Australia Limited (HAL) in 2011 (Comiskey, 2011). HAL was restructured in 2014 into Horticulture Innovation Australia (Hort Innovation), a grower-owned, not-for-profit company and the industry’s declared services body under the Horticulture Marketing and Research and Development Services Act 2000.

Although the banana industry has evolved since 2010, its core production structure has remained relatively stable. The most significant change was the emergence of Panama Tropical Race 4 (TR4) in Far North Queensland in 2015, which led to the establishment of stringent biosecurity zones and protocols (ABGC, 2018; Biosecurity Queensland, 2021). Concurrently, sustainability has become a priority, with initiatives focused on water efficiency, reduced chemical use and integrated pest management (Hort Innovation, 2022). Other developments include greater vertical integration, supply chain consolidation among large producers (Hort Innovation, 2020), and persistent labour shortages exacerbated by the COVID-19 pandemic (ABGC, 2021). Despite these changes, the industry’s fundamental structure and market orientation remain broadly consistent, making the 2011 dataset a suitable reference for this proof-of-concept study.

The original HAL study included 46 banana growers, covering 22% of Australia's banana production area in 2008/2009. Participating farms varied widely in scale, with annual turnovers ranging from under $250,000 to over $10m. After excluding incomplete or outlier cases, data from 27 farms were retained for analysis. The DEA model uses one output and four inputs (Table 4). While the dataset represents a single production year, it is considered sufficient for analysis, as the number of decision-making units (DMUs) exceeds three times the total number of inputs and outputs, providing adequate degrees of freedom for robust efficiency estimation (Cooper et al., 2006).

To determine if the output and the inputs are appropriate for use in the model, correlation between the partial indicators of efficiency were examined to see if they yielded meaningful information. An output-to-input ratio was calculated for each of the inputs, and correlation coefficients between the partial indicators for all farms were then computed (see Table 5). The results show that the different partial indicators can yield consistent yet diverse outcome regarding farm-level performance. They are deemed useful in measuring the performance of a farm from various perspectives to derive an overall efficiency score.

Using a VRS input-oriented model to analyse the data, the efficiency scores of the 27 farms are shown in Table 6. The distribution of the efficiency scores is shown in Table 7.

Taking Farm 1 in Table 6 as an example, the CRSTE of 0.452 suggests that, assuming CRS, the farm could reduce its inputs by 54.8% (i.e. 1–0.452) and still be able to produce the same output. In a CRS model, technical efficiency scores have the same values in an input or an output orientation. It means that Farm 1 can increase its output by 54.8% using the current inputs. In other words, if CRS is assumed, Farm 1 can either reduce its inputs by 54.8% or increase its output by 54.8% to become efficient (i.e. 100%).

As an input oriented VRS model is selected, the focus is on input reduction. The VRSTE of 0.696 implies that Farm 1 has a “pure” efficiency score of 69.6%. The inefficiency is due to poor management. By improving the operation of the farm alone, 30.4% (i.e. 1–0.696) of inputs could be saved. The SE of 0.650 implies a scale efficiency of 65%. The farm is facing increasing returns to scale suggesting that to reach optimal scale the farm could increase its size of operation through internal growth or by merging or collaborating with another farm which is also facing increasing returns to scale. By adjusting the farm to its optimal size alone, 35% (i.e. 1–0.65) of inputs could be saved.

To improve its efficiency, Farm 1 can analyse the practices of Farms 20, 27 and 2, which are identified as its peers or benchmarks. The lambda weight associated with each peer corresponds to its relative importance among the peer group. Ideally, Farm 1 should analyse the best practices from all the three farms. In reality, it could concentrate its best practice analysis on the peer associated with the highest lambda value (i.e. Farm 20).

It should be noted that DEA calculates relative and not absolute efficiency scores. Although firms on the efficient frontier are given a 100% efficiency score, it is likely that they could further improve their productivity through various improvements such as technological upgrades.

Table 7 shows that 11 (41%) out of 27 farms were more than 90% technically efficient. About eight (or 30%) are found to be 100% efficient relative to the others. In terms of scale efficiency, they are not doing too well, as only six (or 22%) were efficient at 90% or higher. All, except two with 100% efficiency, are facing increasing returns to scale, implying that there is room for increasing the farm size to improve efficiency. In other words, while there is scope for many farms to improve their farming practices and farm management to become more technically efficient, there is a much bigger room for increasing their scale of operation by expanding the farm size or collaborating with other farms to achieve economies of scale.

Checking the robustness of a DEA model is essential, as DEA results can be sensitive to data variations, model assumptions and sample size. In this study, we conducted a sensitivity analysis to examine how changes in input and output data affect the efficiency scores. Specifically, we varied the input and output values by ±10% to observe whether the efficiency scores shifted significantly.

As shown in Table 8, only minor changes in efficiency scores were observed when input and output values were adjusted by 10%. The returns to scale for individual farms remained unchanged, and the reference benchmarks or peers were consistent, with Farm 20 as the most efficient performer for others to emulate. These findings indicate that the DEA results are robust in this context and that DEA is a reliable tool for identifying opportunities to improve efficiency and reduce waste at the production stage of the banana supply chain in Australia.

Ensuring the external validity of DEA results is also important, as it determines whether the findings can be generalised to the wider population of similar DMUs. In other words, it tests whether the estimated efficiency scores and benchmarks are likely to hold beyond the specific sample analysed. If the sample is too small, the constructed efficiency frontier may not accurately represent the true production possibility set. This can lead to efficiency scores that reflect sample-specific characteristics rather than actual performance differences, thereby weakening generalisability. A larger and more representative sample, on the other hand, produces a more robust frontier that better captures the diversity of production technologies and practices, increasing the likelihood that the results approximate true efficiency across the broader population (Banker et al., 1984). Therefore, using a sufficiently large sample is necessary to ensure external validity. A common rule of thumb is that the number of DMUs should be at least three times the total number of input and output variables (Coelli et al., 2005). In this study, the sample size is 27, which exceeds the minimum required number of 15 (based on five input and output variables). Hence, external validity should not be a concern in this case.

To evaluate whether the weights assigned to various input and output variables significantly impact DEA outcomes, entropy-based weights were applied as described in Section 3.2. Table 9 presents the entropy-based weights calculated using proportional scaling and min-max normalisation.

With the min-max normalisation method, higher weights are assigned to Input 3 (packaging cost) and Input 1 (wages and contract labour services), highlighting their relative importance. These higher weights indicate low variability or uncertainty (i.e. low entropy), suggesting that the data for these input variables are more uniform or concentrated, making them more informative. Consequently, these variables play a more significant role in distinguishing between alternatives. In contrast, proportional scaling does not provide such insights when calculating weights.

Running the VRS model with the weighted input and output values gives similar findings, demonstrating the robustness of the DEA approach. A comparison of the traditional DEA, the proportional scaling entropy-based DEA and the min-max normalisation entropy-based DEA is shown in Table 10.

The findings of the entropy-based DEA align closely with those of the traditional DEA, particularly when using weighted values calculated through proportional scaling. This indicates that proportional scaling does not enhance the discriminative power of the DEA model, as it fails to effectively account for the variability or uncertainty in the dataset. In contrast, the use of the min-max normalisation approach for weight calculation has shown an improvement in the model's discriminative power, albeit with a limited effect in this case.

As shown in Table 10, the efficiency scores derived from the entropy-based DEA using min-max normalisation are more concentrated compared to the other two models. Furthermore, it reveals that the standard deviation and range of VRSTE (MMN) are smaller, and the top benchmarks for most farms are more consistent. This improved discriminative power enhances the DEA’s effectiveness as a decision-making tool, enabling government and grower associations to develop more targeted strategies and action plans to improve efficiency and sustainability in the industry.

Table 11 presents the average input savings across all 27 farms using the entropy-based DEA. Achieving 100% technical efficiency through improved farm management could result in average input savings of 26.9% in employee wages, 34.7% in fertiliser use, 21.7% in packaging costs and 33.1% in on-costs. By comparison, the average savings calculated with traditional DEA are 25.3, 34.7, 20.6 and 30.0%, respectively. While the differences are minor, they highlight the greater discriminatory power of the entropy-based approach. It is important to note that in a VRS model, improvements in variables (reductions in inputs or increases in outputs) are based solely on the VRS technical efficiency score, without accounting for scale efficiency. This indicates that additional savings could be achieved through scale improvements alongside farm management enhancements.

Table 12 highlights that, on average, the most efficient farms outperformed their less efficient counterparts by producing 36.8% more output while using 21.6% less labour and 21.5% less fertiliser. These farms also demonstrated greater cost efficiency, with packaging and farm management costs reduced by 5.5 and 20.6%, respectively. In contrast, the differences calculated using the traditional DEA method were 37.7%, −19.6%, −24.2%, −7.4% and −12.0%, respectively. The larger disparities arise from the traditional DEA's lower discriminatory power, which classified eight farms as 100% technically efficient compared to seven classified by the entropy-based DEA. Consequently, the average performance of these technically efficient farms appears higher when assessed using the traditional DEA.

In this section, a strategic discussion is conducted to interpret the key findings in the broader context of the Australian banana supply chain, highlighting the managerial implications, providing the appropriate strategic solutions and projecting the likely outcomes. This study found that most farms are not operating at top efficiency level for various reasons, including deficient farming practices, defective farm management or suboptimal scale. Some farms are 100% efficient on the efficiency frontier, while others are less efficient. The gap between the CRS and VRS frontiers highlights scale issues. The CRS model assumes farms operate at optimal scale in a perfectly competitive environment, which is rather rare in practice. The VRS model, however, accounts for variable returns to scale, recognising that farms may face economies or diseconomies of scale due to constraints. There is a clear and significant difference in performance between the most efficient farms and those which are less efficient. There are farms which are technically efficient under the VRS model but not fully efficient under the CRS model. This implies a scale inefficiency.

The HAL’s study (Comiskey, 2011) on Banana Enterprise Performance Comparison also highlighted significant differences in the farm management activities of individual growers, which in turn could lead to a wide variation in business profitability. In that study, average inputs and outputs of the top 10 farms were compared with the rest to identify the performance gaps. As the comparison was based on one variable at a time and the top ten farms (out of 46 growers) differed in each comparison, the outcome was not as consistent as that of the current study using DEA. Nevertheless, it still provided a general picture which could be used as a cross-validation of the current DEA approach. The HAL’s study (Comiskey, 2011) findings showed that the top ten farms were 39.6% more productive (in terms of cartons per hectare) than the remainder. It was reported that a primary reason for this could be the strong crop management skills of these growers, as they had the ability to produce significantly larger percentages of extra-large bananas than other growers. Factors of success may include economies of scale, access to technology and infrastructure, professional management and expertise, stronger supply chain integration, better risk management, continuous learning and innovation and relatively higher labour stability. The top ten farms also had labour costs (owners, employees and contractors) 25.5% lower per carton. This could be indicative of these growers having more developed human resources management skills and more efficient farming and packaging systems. Fertiliser and chemical usage costs were 9.8% lower than for the remaining growers. The lower costs per carton were believed to be largely due to the higher yield per hectare. The on-farm costs per carton were 25.9% lower, suggesting that the farm management practices of the top ten farms could be more cost-effective. The DEA approach adopted in this study provided a more comprehensive and consistent analysis of the efficiencies of the farm to reach the same conclusion. Not only did it highlight for each farm the areas for improvement in terms of input utilisation but also identified the more efficient peers to learn from in terms of farm management practices. As such, the strong capability of DEA as a benchmarking tool was clearly demonstrated.

Whilst there are significant differences noted in farm-level efficiency, these differences are also influenced by farm location and distance from markets, production scale, enterprise sophistication and resources, management preferences and contractual arrangements. For example, on-farm food waste varies between 10 and 30% in North Queensland. The strategies and actions necessary to improve farm efficiency and reduce waste across the banana supply chain need multi-agency coordination and supply chain collaboration driven by lead growers. The Australian banana supply chain is largely contained within Australia, with no fresh bananas imported and only a very small percentage exported. This differs from many other countries where there are substantial investments in the trade of bananas to manage the risk of a more complex banana supply chain. For example, the US banana imports totalled 5.08 m metric tons in 2023 (Karst, 2024).

The DEA results of this study provide a quantitative basis for identifying resource inefficiencies and scale-related performance gaps among banana growers, thereby guiding targeted strategic interventions. Farms with low technical efficiency but operating at an optimal scale may benefit from benchmarking and best practice sharing to improve input utilisation and operational management (Coelli et al., 2005). In contrast, farms exhibiting scale inefficiency, particularly those operating under increasing returns to scale, indicate that they are too small to exploit economies of scale efficiently (Tone, 2001). For these growers, strategies such as cooperative models, resource sharing or vertical integration can collectively enhance input purchasing power, access to technology and market coordination, enabling them to achieve scale efficiency comparable to larger farms (Bijman and Iliopoulos, 2014). Conversely, large farms operating under decreasing returns to scale may need to rationalise resources or diversify downstream activities to mitigate inefficiencies and reduce surplus-driven waste. Thus, the DEA results not only benchmark individual farm performance but also provide evidence-based justification for strategic pathways tailored to farm size and efficiency profile, linking micro-level inefficiencies to meso-level structural solutions for improving resource efficiency and reducing food loss across the banana supply chain.

Table 13 presents the key findings of this analysis with their managerial insights, coupled with recommended supply chain strategies to generate desirable outcomes. The supply chain strategies for farm efficiency improvement include a closer alignment between production and customer demand, supporting the adoption of industry best practice including the use of cold food chain, dealing with cosmetic imperfections and a sustained supply of skilled labour in peak seasons. Implementing key management models, such as total quality management with zero defects, Six Sigma or Muda to reduce waste and Kaizen with small incremental changes for continuous improvement, that espouse monitoring evaluation reporting and improvement may help reduce waste and improve efficiency and sustainability of production, processing and distribution of the banana supply chain. However, it is contended that more specific and targeted strategies would be more effective strategies to help develop or improve, assess, implement and review and reflect on risk factors affecting the efficiency of the banana supply chain in Australia. We summarise and discuss this table using four broad supply chain strategies in tackling the issues driving farm-level inefficiency as evidenced in this study.

Inter-farm comparisons reveal that larger, corporatised farms with technological innovations tend to operate more efficiently, particularly in achieving economies of scale. The farm-level differences should therefore be distinguished between larger banana growers and their facilities and customers and the medium and smaller farmers who rely more importantly on industry intermediaries. The average farm size in Australia is about 400 hectares, ranging from 50 hectares to thousands of hectares. With size comes different modes of planting, harvesting and processing, with the bigger farms able to employ economies of scale to use more advanced technology and automation. In addition, product processing after harvesting of bananas involves quality inspection, sorting, washing and packaging. Bigger farms have these facilities on site, whereas smaller farmers rely on dedicated packing facilities to process their product.

The farms can improve by adjusting its size, either through splitting into smaller units or merging with others. Farms below the VRS efficiency frontier must first address managerial inefficiencies to become VRS efficient. Only then they can tackle scale inefficiencies to achieve CRS efficiency. Farm-level efficiency is important because crop waste for bananas in Australia was 29% per farm in 2021–2022 and 24% in 2022–2023, respectively, according to the ABARES surveys (Downham, 2022; Slatter, 2023). Arguably, a horizontal integration strategy is recommended to address various sources of inefficiencies among smaller farms. This approach involves consolidating smaller landholdings into larger operations, enabling larger banana growers to expand their plantations through acquisitions or mergers with smaller farmers. Such consolidation can increase market share, reduce competition, achieve economies of scale and diversify product offerings. Cost efficiency improvement through forward or backward vertical integration by larger retailers is likely to play a significant role in the banana production industry. This integration will enable resource sharing, stronger bargaining power and streamlined operations. Banana farms in Queensland, New South Wales and South Australia exhibit notable variations in farm-level efficiency. To capitalise on differences in labour costs, land value, and land productivity across these regions, geographic integration presents a viable strategy for larger banana growers. Expanding supply chains into new regions allows growers to optimise resources, tap into broader markets and enhance operational efficiency by leveraging areas with lower production and logistical costs.

Collaborative strategies can effectively address inefficiencies in packaging, storage and long-haul transportation of bananas. For instance, small farmers might transport their produce directly from the field to a nearby cooperative packing shed or pack it on-site before consolidating it at a local warehouse with other growers’ products for further transportation or ripening. Bananas are transported usually by road with climate-controlled trucks that act as temporary storage while travelling to allow ripening. Since the product is delivered Australia-wide, the distances covered and time taken vary according to the final destination of consumption. These initial distribution points can be major supermarket chains or state markets in the capital cities. These again provide warehousing or storage while ripening and preparing for. Some growers bypass the central market system altogether by establishing contracts with major supermarkets, enabling direct delivery to retailer warehouses or distribution centres.

Cooperative bulk purchasing is another valuable strategy, allowing small farmers to partner with others to procure packaging materials in bulk at reduced costs. This approach also empowers them to secure long-term contracts with suppliers for volume discounts and stable pricing. Scaling operations can improve palletisation and stacking methods, maximising shipment capacity and reducing per-unit packaging costs. Additionally, direct-to-retail packaging, i.e. designing packaging that doubles as retail display units, can minimise handling and eliminate the need for extra packaging at the destination, further enhancing efficiency.

Forming a cooperative enables small farmers to consolidate their produce and share logistics services to achieve efficiency and resource sharing. Some farmers, for example, may choose to contract or subcontract their produce or services to other actors in the chain, to form cooperatives to be able to enjoy economies of scale or to engage in becoming buffer suppliers to compensate for the vagaries of supply and demand experienced by other producers, e.g. contractual obligations of the big producers to the big supermarkets. The expanded supply chain provides an opportunity for farmers to improve their current operations through options such as collaboration, co-opetition, forming cooperatives and contracting or subcontracting parts of their product or services.

Labour shortages are a persistent challenge in regional and rural Australia, often addressed through programmes like the Seasonal Worker Programme (SWP) and Pacific Australia Labour Mobility (PALM) as well as employing backpackers. To optimise labour use, growers should implement accurate demand forecasting to prevent overstaffing or understaffing during peak and off-peak seasons. Training farm workers for multiple roles can enhance flexibility and minimise excess labour during downtime.

Larger growers can reduce reliance on third-party contractors by directly recruiting and managing their workforce, providing better control over wages, working conditions and alignment with productivity goals. Contracts should be tied to productivity and quality standards, with incentives to promote efficiency while upholding worker rights.

One effective strategy is output-based pay, where workers are compensated per unit of work (e.g. per box of bananas picked) using fair piece-rate systems. These systems should ensure transparency, compliance with minimum wage laws and alignment with realistic productivity expectations. However, this approach must carefully address potential risks, such as worker exploitation, inconsistent earnings, quality control issues and worker’s physical and mental health.

To help operationalise the strategies (Section 5.1–5.4) to reduce banana waste across the broader supply chain, a ten-point action plan developed around four domains, namely infrastructure and logistics, procurement and contracts, packaging practices and retail engagement, is proposed (Table 14).

This study shows that the sustainability of banana farms can be improved by comparing their operating efficiencies and identifying areas for improvement using the DEA technique. The findings can help reduce wastage at the growing and on-farm (harvesting, grading and packing) stage, which is one of the major issues in the horticulture industry. Results from the entropy-based DEA show that among the 27 farms in the states of Queensland, New South Wales and South Australia, 37% are found to have a technical efficiency score over 90%. About 26% are found to be 100% efficient relative to the others. Two farms are found as major peers or benchmarks for others to learn from. Only 22% of farms are found to have scale efficiency scores over 90%. All, except two with 100% efficiency, are facing increasing returns to scale, implying that there is room for increasing the farm size to improve efficiency. The average savings upon improvement in farm management (i.e. to become 100% technically efficient) in the four inputs are found to be 26.9% in employee wages, 34.7% in fertiliser usage, 21.7% in packaging cost and 33.1% in on-costs, respectively. The findings align with the traditional manual-based category comparison approach, but the DEA approach can provide more objective results with greater discriminative power to facilitate strategy formulation for efficiency improvement. Building on the findings of this study, future research on the banana supply chain should explore a broader range of variables that capture the evolving nature of farm management practices and technological adoption in horticulture. While this research has demonstrated that improved utilisation of labour, fertiliser, packaging and other resources contributes to reducing production costs and enhancing sustainability, it is important to acknowledge that farming practices and technologies have advanced significantly in recent years.

Future investigations should examine the role of emerging technologies such as precision farming, the Internet of Things (IoT) devices, artificial intelligence (AI) and tracking and tracing systems in improving farm efficiency and minimising waste. Specifically, there is a need to identify and include additional input and output variables that reflect the technological impact on farm performance. For example, smart sensors that monitor and optimise the use of water and electricity can be critical indicators of efficiency, particularly in the context of environmentally sustainable agricultural practices.

Moreover, spatial and temporal factors should be integrated into future analytical models. Geographical attributes such as the latitude and longitude of farms, soil characteristics and climatic conditions over time (e.g. temperature and rainfall patterns) can significantly influence productivity and resource use efficiency. Incorporating these variables into models such as DEA will enable more accurate benchmarking and targeted recommendations for waste reduction.

Finally, the effectiveness of advanced efficiency measurement tools like DEA is highly dependent on the availability and quality of farm-level data. However, such data are often not systematically collected, may be outdated or remain inaccessible due to privacy or policy constraints. Future research should therefore also focus on strategies to improve data collection frameworks and encourage data sharing practices, potentially through government or industry-led initiatives.

By addressing these research gaps, future studies can provide deeper insights into the multifaceted drivers of horticultural waste and offer more tailored strategies for achieving long-term sustainability in the banana supply chain.

Efficiency can be assessed through three key metrics: technical efficiency (TE), pure technical efficiency (PTE) and scale efficiency (SE) (Farrell, 1957; Le et al., 2022). TE reflects a unit’s ability to produce the maximum output from a given set of inputs. PTE isolates the impact of managerial performance by removing the effects of scale, while SE evaluates whether a unit is operating at its optimal production size. Scale efficiency is typically measured as the ratio of TE to PTE. Farrell (1957) also introduced the concept of allocative efficiency (AE), which focuses on the cost-effective mix of inputs based on their relative prices. However, AE requires input price data, which is often unavailable in agricultural settings, making TE and SE more practical measures in many studies.

DEA, introduced by Charnes et al. (1978), is one of the most widely used non-parametric methods for measuring efficiency. It evaluates DMUs, such as farms or firms, by comparing the ratio of weighted outputs to weighted inputs, using linear programming to construct an efficiency frontier based on the most efficient performers (Thanassoulis et al., 2008; Cooper et al., 2007). DEA models can be input-oriented, aiming to minimise inputs while maintaining output levels or output-oriented, focusing on maximising outputs for a given level of inputs (Paul, 2024; Paul et al., 2017).

One of DEA’s strengths is its ability to handle multiple inputs and outputs without requiring a predefined production function, reducing the risk of model misspecification (Košařová and Pokrivčák, 2023). This flexibility makes DEA particularly valuable for evaluating farms or production systems with heterogeneous practices and diverse outputs, where conventional parametric models may fall short (Atici and Podinovski, 2015; Picazo-Tadeo et al., 2011; Wagan et al., 2018).

Several DEA models have been developed to address different analytical needs. The CCR model assumes constant returns to scale (CRS) and measures overall efficiency, while the BCC model allows for variable returns to scale (VRS) and isolates managerial (purely technical) efficiency (Banker et al., 1984; Kyrgiakos et al., 2023). Although these models are relatively simple and require only quantitative input–output data, they may struggle to differentiate among efficient DMUs or account for contextual differences (Hong and Jeong, 2020; Le et al., 2022).

To overcome these limitations, more advanced and context-specific DEA models have been introduced. These include models that incorporate undesirable outputs, account for weight restrictions or integrate external variables. Such developments enhance the discriminatory power and relevance of DEA for diverse applications in sustainability, agriculture and supply chain efficiency analysis. Table A1 summarises the main DEA models, their applications and their relative strengths and weaknesses.

The gap between the CRS and VRS frontiers in a DEA model reveals scale inefficiency. A firm with a VRS technical efficiency (VRSTE) score of 1 is technically efficient relative to firms of similar scale. However, if its CRS technical efficiency (CRSTE) score is less than 1, it indicates suboptimal scale. The firm could improve overall efficiency by adjusting its size, either expanding (e.g. through mergers) or downsizing (e.g. splitting into smaller units), to better align with the CRS frontier.

If a firm lies below the VRS frontier, it suffers from both managerial inefficiency and scale inefficiency. The appropriate course is to first improve management practices to become VRS-efficient. Once this is achieved, the firm can then focus on adjusting its scale to reach full CRS efficiency.

DEA models can be input-oriented or output-oriented, depending on which variables the decision-maker controls more directly. An input-oriented model seeks to minimise input for a given level of output, while an output-oriented model aims to maximise output for a given level of input. Although the CRS model yields the same efficiency score regardless of orientation, VRS model scores can differ due to scale effects.

Returns to scale provide further insight. A firm operating under increasing returns to scale can achieve greater efficiency by expanding its operations, as output increases faster than input. Conversely, a firm under decreasing returns to scale may benefit from downsizing, as further expansion leads to disproportionately higher input use. A totally efficient firm faces constant returns to scale, meaning it is operating at its optimal size, where inputs and outputs increase proportionally and average productivity is maximised.

DEA also identifies “peer” firms on the efficiency frontier as benchmarks for underperforming firms. These efficient peers offer best practices that can guide improvement, helping inefficient firms to close performance gaps and move closer to the frontier.

The mathematical formulation of the VRS DEA model with input orientation employed in this research is as follows:

Objective

$minθ,λθ$

Subject to

Definitions:

θ: Efficiency score of the DMU being evaluated (θ ≤ 1).

λ_j: Weight assigned to DMU j.

x_ij: Input i used by DMU j.

y_rj: Output r used by DMU j.

x_i0: Input i used by the target DMU (DMU₀).

y_r0: Output r used by the target DMU (DMU₀).

n: Number of DMUs.

m: Number of inputs.

s: Number of outputs.

1. The objective function minimises the input contraction factor θ, which represents the proportion by which all inputs can be reduced while maintaining output levels.

2. Constraint (1) ensures that the scaled inputs of the target DMU are at least as large as the weighted sum of inputs for all DMUs.

3. Constraint (2) ensures that the outputs of the target DMU are at least as large as the weighted sum of outputs for all DMUs.

4. Constraint (3) enforces convexity, making it a VRS model.

5. Constraint (4) ensures non-negativity of the weights λ_j, as negative weights are not meaningful in this context.

The steps to calculate the weights of the input and output variables of the DEA model based on Shannon's (1948) entropy-based weighting mechanism are as follows:

• Step 1: Problem setup

Suppose there are n DMUs with:

Inputs: x_ij (where i is the input index and j is the DMU index)

Outputs: y_rj (where r is the output index and j is the DMU index)

• Step 2: Normalise input and output values

To make all the values comparable, normalise input and output values as z_ij and w_rj using the following formula:

$zij=xij∑j=1nxij$ and $wrj=yrj∑j=1nyrj$

• Step 3: Calculate entropy for each input and output

The entropy E_i for each normalised variable is calculated using the following formula:

$Ei=−k∑j=1nzijln⁡(zij)$ and $Er=−k∑j=1nwrjln⁡(wrj)$⁠, where $k=1ln⁡(n)$⁠, n = number of DMUs

• Step 4: Calculate entropy-based weights

The weight for each variable is calculated using the following formula:

$wi=1−Ei∑i=1m(1−Ei)$ and $wr=1−Er∑i=1m(1−Er)$

• Step 5: Apply DEA with entropy-based weights

Run the traditional DEA models with the weighted inputs and outputs to determine the efficiency scores of the DMUs.

Inputs: z_ij × w_i

Outputs: w_rj × w_r

In Step 2 above, the sum method is used to normalise input and output values. This method produces values that represent the proportion of each individual value relative to the total. It does not transform values to a fixed range. The sum method is frequently called “proportional scaling” or “normalisation by sum” but is not considered “normalisation” in the strict mathematical sense. As such, the min-max method is often used instead (Roszkowska and Wachowicz, 2024). This method scales values proportionally based on their range and transforms the values of a variable to a specified range, usually [0, 1] or [−1, 1]. It retains relative relationships (ratios) between data points and is commonly referred to as true “normalisation”. The following formula is used:

$zij=maxixij−xijmaxixij−minixij$ and $wrj=yrj−miniyrjmaxiyrj−miniyrj$

For the purpose of comparison, both the proportional scaling and the min-max normalisation method are used in calculating the entropy-based weights for the DEA model.

This paper forms part of a special section “Advancing Sustainable Connectivity: Innovations in Logistics and Supply Chain Integration (ISL 2024)”, guest edited by Ruth Banomyong and Kamrul Ahsan.