2.1. Effects (input KPIs)

Human-centric factors related to manual tasks can be summarised by two key questions: 1) are they working? 2) if so, how well are they working? (Reference Salzer, Naor, Yonai, Yechezkely, Kfir, Zait, Dionysis, Charisios, Georgios and MariaSalzer et al., 2021). We selected these two questions because they capture the most critical aspects of manual labour that directly affect productivity. The first question is addressed through Availability and Attendance , in which availability reflects whether workers are accessible for tasks, while Attendance indicates their actual presence during work. These parameters are essential because worker presence is necessary for productivity (Reference Barbulescu, Vargas-Silva and RobertsonBarbulescu et al., 2021). The second question is more complex, as worker performance is influenced by experience, physical conditions, and other factors. To capture these influences in a quantifiable way, we introduced the parameter Capability which combines work speed and decision-making accuracy (Reference Bedford and BedfordBedford & Bedford, 2013).

• Availability is implemented by adjusting the total number of worker agents generated on model initialisation, which is determined by scenario-specific probabilities.

• Attendance is defined as the percentage of worker agents who would work during a day.

• Capability is understood as 1) the speed of the worker performing the tasks; and 2) the quality of the product, which is the quality of the harvested product. Additionally, there is the following relationship between these two:

(1) $$R_{{{high}}} = {{dP_{{{high}}} } \over {dt}} = {{dP_{{{tot}}} } \over {dt}}J_{{{high}}}^p $$

And

(2) $$R_{{{low}}} = {{dP_{{{low}}} } \over {dt}} = {{dP_{{{tot}}} } \over {dt}} - {{dP_{{{high}}} } \over {dt}}$$

Where R _high and R _low are the rate of picking high-quality and low-quality apples, P _high and P _low are the number of high-quality and low-quality products one single harvested, P _tot is the total number of products one single harvested, and $J^p_{high} $ is the assessing criteria for each product being high-quality.

2.2. Worker behaviours

In agricultural production tasks that are manual, worker behaviours are the core factor affecting production efficiency and output quality. This paper further divides “Human Factors“ into the following four main categories:

• Harvest is the core task of the entire harvest process which is reflected by the picking rate. The harvest of each worker during the picking process can be calculated as follows:

  (3) $$P_{{{tot}}} = (R_{{{high}}} + R_{{{low}}} )*t_{{{harvest}}} $$

  Where t _harvest is the time of a single picking.

• Transportation represents the process of the harvested products transported to the designated location to complete the collection, which can be calculated as follows:

  (4) $$t_{{{trans}}} = {{L_{{{path}}} } \over {v*\mu }}$$
  Where t _trans is the transportation time of a single picking, L _path is the length from the tree to the designated location, ν is the worker transportation speed, and μ is an efficiency factor related to physical condition and tool use.

• Judging is the key behaviour in the harvest process which can be reflected by the “judgement accuracy” parameter. This involves workers assessing criteria such as the quality of the product, namely $J^p_{high} $ .

• Self-growth shows an improvement in worker abilities in multiple tasks, that is, the “learning effect”. This is primarily demonstrated through an increase in picking speed and judgement accuracy over a specific time and picking counts, highlighting the workers’ capacity to adapt and enhance their efficiency. The following linear growth equation is used here:

  (5) $$X = X_0 *(1 + G_X )$$
  Where X is the improvement parameter, namely $J^p_{high} $ and R _high , X ₀ original value, and G _X is the self-growth factor.

2.3. Impacts (output KPIs)

Impacts on the agricultural sector were described in terms of yield, production, or rate of these. Rates were given per person, per time or both, for example ‘per man-hour’. To quantify the impacts of changing input parameters, the following output KPIs were proposed for framework performance.

• Yield represents the total number of products harvested during a simulation run and can be calculated as:

  (6) $$P_{{{tot}}}^p = P_{{{high}}}^p + P_{{{low}}}^p $$
  Where $P^p_{high} $ $P^p_{low} $ and $P^p_{tot} $ are the number of high-quality, low-quality, and the total number in storage at the end of the simulation.

• Production represents the total value of the products harvested during a simulation run, which can be calculated by the following equation:

  (7) $${{Production}} = {{Weight}}_{{{low}}} *{{Price}}_{{{low}}} *P_{{{low}}}^p + {{Weight}}_{{{high}}} *{{Price}}_{{{high}}} + P_{{{high}}}^p $$
  Where Weight _low and Price _low are the average weight and price per weight of low-quality apples, Weight _high and Price _high are the average weight and price per weight of high-quality apples.

• Efficiency and Productivity are measures of yield per worker-hour and production per worker-hour, respectively. They are calculated as

  (8) $${{Efficiency}} = {{P_{{{tot}}}^p } \over {N_{{{tot}}} T}},{{Productivity}} = {{{{Production}}} \over {N_{{{tot}}} T}}$$

Where N_tot is the total number of workers available in a simulation and T is the time over which a simulation is run measured in hours.

3.1. Orchard environment

Given this framework, relevant parameters were considered to create an interactive orchard environment to establish a reasonable distribution of apple trees to meet the picking and transportation needs of the picker agents. The following three parameters were considered:

• Apple Type (Gala): Different varieties of apples are suited to different growing conditions and therefore different orchard properties. Due to the popularity of Gala apples, which were named the most-produced apple variety in the UK by British Apple and Pears Limited, Gala apples were selected as the variety choice for the simulation (British Apples & Pears Ltd, 2023).

• Tree Spacing (4.5m × 6m): The optimum tree density for modern orchards is close to 1000 trees/acre (405 trees/hectare) (Reference Robinson, Hoying, Sazo, Marree and DominguezRobinson et al., 2013). This aligns with the MM111 dwarfing rootstock tree spacing of 4.5m between trees in a row and 6m between rows, yielding an overall orchard density of 370 trees/hectare (The Royal Horticultural Society, 2024).

• Orchard Size (143.5m × 160m): An efficient orchard should have a maximum row length of 152m and enough unplanted space left at each end of the orchard for the turning off equipment (Western Agricultural Research Center, 2024). Therefore, an orchard consisting of 25 rows and 30 trees per row, with 4m unplanted at each orchard end was built, which resulted in an overall orchard size of 143.5m × 160m.

Figure 2 shows an example of the orchard environment with reduced tree count simulated in Anylogic. Along with trees, the environment is populated with three other agent types: pickers, bins, and a tractor. As apple orchards are planted in single columns with alleys wide enough for a tractor on either side, pickers also move along alleys and only pick from one-half of the tree at a time. To reflect this, trees were modelled as halves belonging to a left or right population. Pickers were then also assigned to a left or right population with members of each only able to pick from the corresponding tree half population. This necessitated the creation of the picker movement networks (green and blue lines) allowing for picker movement along alleys but between them only at the ends of tree rows.

Figure 2. The simulated orchard environment using a reduced tree count for visibility

3.2. Picker decision making

Further following the autonomous ABM approach, pickers were modelled as individual agents with internal behaviour dictated by a discrete event state chart that allows them to autonomously interact with the orchard environment. The picker decision-making process is detailed (Figure 3) which was expanded from the primary picker task order: Determine the ripeness of the apples, then pick qualified apples and put them in buckets/bags (At tree), walk to central container/bin, empty the full bucket/bag (At bin), walk back to the tree and repeat, and as the number of pickings increases, the picker will achieve self-growth and thus improve the picking speed (Reference Salzer, Naor, Yonai, Yechezkely, Kfir, Zait, Dionysis, Charisios, Georgios and MariaSalzer et al., 2021).

Figure 3. Flow diagram representing the discrete event logic implemented as the picker decision-making process

3.3. Simulation parameters settings

For picker logic to be accurately implemented, the following physical picking parameters were derived and set as model parameters.

• Apple Weight and Price: The weight and price of Gala apples are assumed based on data from the literature, and there are corresponding differences due to differences in quality (Reference Hall, Tustin, Silva and StanleyHall et al., 2015; Ontario apple growers, 2016). The average weight of high-quality apples is 200g and the price is 0.881$/kg, while the average weight of low-quality apples is 166g and the price is 0.302$/kg.

• Pick Rate (73.3gram/person/second): The apple picking speed is assumed to be completed on a fixed platform, which is calculated based on the average throughput (Reference Fei and VougioukasFei & Vougioukas, 2021). Additionally, the action cost of climbing (5.4%) and moving (6.3%) the platform is added to the picking speed (Reference ZhangZhang, 2015).

• Worker transportation speed (1.25m/s): In this model, it is assumed that the transportation is done by human labour. Considering the weight of apples, the transportation speed is set to 1.25m/s which is a 20% reduction from normal speed (Reference BohannonBohannon, 1997). The simulation environment determines the path length in Equation 4, and the efficiency of all workers is assumed to be 1.

• Bucket Capacity (17kg): Apple-picking buckets were designed to hold 17 kg of apples when full. This capacity was adopted in our simulation to ensure alignment with empirical data and to maintain consistency in modelling manual harvesting scenarios (Reference Freivalds, Park, Lee, Earle-Richardson, Mason and MayFreivalds et al., 2006).

• Bin Capacity (400kg): A typical apple bin has a volume of approximately 0.71 m³ (Reference Freivalds, Park, Lee, Earle-Richardson, Mason and MayFreivalds et al., 2006). Assuming an average apple bushel weighs 45 lbs, this results in a bin capacity of around 400 kg or approximately 2,000 apples (Reference Myers, Corditz, Graberry and KelleyMyers et al., 2022).

• Tree Yield (23.7kg/tree): Figure 4 was used to predict the orchard yield for the chosen orchard density and therefore the per-tree season yield of 47.4 kg/tree. Gala apples are picked over multiple harvests, with the first pick harvesting half the apples available apples (Reference Hall, Tustin, Silva and StanleyHall et al., 2015). Using this assumption, the model environment simulated a first pick by assigning a tree yield of half the actual tree yield at 23.7 kg/tree.

Figure 4. Orchard yield plotted against the density of trees (Treder & Mika, 2001; Hampson et al., Reference Hampson, Quamme and Brownlee2002; Robinson et al., Reference Robinson, Hoying, Sazo, Marree and Dominguez2013; Ontario apple growers, 2015; Reference Treder and MikaMinnesota Fruit Research, 2018)

3.4. Scenario settings

Having established a simplified model capable of simulating human activities in orchard harvesting, this section demonstrates this model in two scenarios based on labour dynamics: the labour shortages in the UK FSC.

3.4.3. Capability

Due to the lack of specific data at the individual level, the differences between pickers are reflected by selecting random numbers within a reasonable range, which ultimately reflects the average value and trend of the model parameters.

• Picking Rate: Pickers’ efficiency fluctuates by ±10% around the average, creating a maximum efficiency difference of 20% between skilled and ordinary workers (Reference Zhao, Binks, Kruger, Xia and StenekesZhao et al., 2018). Considering the decline in picker quality, the picking rate under labour shortages is 5% lower than the original.

• Judge Factor: When considering the quality of choice in apples, the average percentage of high-quality apples is 87% (British Apples & Pears, 2023). Consequently, the average judgment accuracy for high-quality apples is set to 0.87, with fluctuations consistent with the picking rate.

• Self-growth: After completing the harvest of 100 buckets or having systematic training programmes, the picker’s picking rate improves by 5% (natural growth) or 10% (training growth), capped at a maximum rate of 80.63 g/person/second. The judgment accuracy also increases, reaching a maximum of 95.7%.

The quantitative modelling of labour availability, attendance, and capability differences clearly reveals the combined impact of labour shortages on the ratio of seasonal and domestic pickers, overall labour reduction, and human factors on picking efficiency and apple quality (Table 1).

Table 1. Human factors parameter design under labour shortages

4.1. Results

Based on the settings in Section 3, the simulation results in two cases are analysed. The first is to simulate picking activities over 3 hours with 10 repetitions to analyse the impact of labour shortages on agricultural production; The second is to simulate production activities within 6 hours and add a training programme for pickers post-shortage in the third hour (i.e., increase the picking speed) to analyse the impact on the entire production. The simulation results presented in Figure 5 and Table 2 offer a comparative analysis of orchard harvesting performance pre- and post-shortage.

The first case results from the 10 simulations reveal a natural variability in the number of apples picked, due to the randomness in picker capabilities (Figure 5 (a)). Despite these fluctuations, all simulations remain within 5% of the average, indicating that while individual performance can vary, the overall trend remains consistent. Table 2 details the changes in output KPIs based on the simulated average picking number between the pre- and post-shortage scenarios used in this simulation. There was a noticeable decline in the total number of apples picked post-shortage, with a 17.3% reduction in high-quality apples and an 18.7% increase in low-quality apples, which also caused a decrease in production by 15.5%. However, the fluctuation of efficiency and productivity is within 5%, which is consistent with the difference in pickers’ capacity in simulation.

This suggests that the shift in workforce composition, with fewer seasonal workers and increased reliance on domestic workers, may have impacted the efficiency and accuracy of picking. Seasonal workers are typically more experienced, and more adept at selecting high-quality fruit, whereas domestic workers may not have the same level of expertise, resulting in a decrease in quality.

The second case results show how picking activity changes over time after additional training (Figure 5 (b)). This process can be divided into three regions, of which 0-1.5 hours is the stable region, which is caused by the lag of the collection tractor at the beginning of picking; 1.5-3 hours is the gap region, at this time, the advantages of more workers and higher skills pre-shortage gradually emerge, and production greatly exceeds that post-shortage; 3-8 hours is the catch region, which means after suitable training, the skill advantages of post-shortage workers gradually emerge, approaching and eventually catching up with pre-shortage production capacity.

This finding underscores the critical role that workforce composition plays in orchard harvesting. Strategies to mitigate the effects of labour shortages, such as better training for domestic workers or the use of technology (e.g., automated picking tools), could be explored to compensate for the reduction in experienced seasonal workers. Additionally, the possibility of improving the selection and training processes for workers could help bridge the gap in picking efficiency and quality.

Table 2. Average output KPI values for pre- and post-shortage simulations

Figure 5. (a) The number of apples picked changes in ten simulations, (b) Production changes over time after additional training

While the post-shortage scenario resulted in a reduction in overall production, the modelling suggests that workforce composition, training, and skill development will be key factors in mitigating these challenges and maintaining productivity levels.

4.2. Discussion and applications

This model adopts a systems approach to simulate the complex interplay between societal challenges and the resilience of food supply chains, with a particular focus on human-centric factors such as labour dynamics, policy impacts, and worker well-being. By adjusting the model parameters (such as the number of labour forces, skill levels, social events, etc.), the model is not only adaptable to various scenarios, ranging from the effects of a global pandemic on labour availability to the impact of social changes but also scalable to different food production systems. Further, while the model is applied to food production, it could be applied to the manufacture of physical goods and services, such as those provided by design agencies to understand how external factors and interventions might affect overall capability. This could include, for example, market, training or technical (design-led) interventions, such as new tools/devices or planning strategies, as well as evaluating potential collaborations with external partners. These insights contribute to sustainability and resilience by enabling the optimization of sustainable labour practices, while also addressing aspects of health and well-being through interventions like skill training and workforce management, which bridges societal challenges with actionable strategies for resilient and efficient systems.