Introducing digital tools for sustainable food supply management: Tackling food loss and waste in industrial canteens

Reducing food waste is essential if we want to create a more sustainable food system, which is why halving per capita food waste by 2030 has been included in the UN Sustainable Development Goals. The main aim of this paper is to demonstrate that by using digital tools to monitor food provisioning and management within canteens, it is possible to achieve a more sustainable food service management within industrial companies and to reduce the overall quantity and environmental impacts of food preparation and consumption thus achieving economic benefits. Longitudinal 2018 and 2019 data collected from the canteen of a major Italian food production company were analyzed. The results showed that the amount of food lost and wasted has decreased over time, reaping important environmental benefits and illustrating specific differences between the food lost in meal preparation and the food left on employees’ plates. It is therefore important to implement education initiatives and to use digital tools to share food‐related data collected within companies among both employees and kitchen and catering staff in order to raise awareness of their behavioral strengths and weaknesses. From this perspective, new interventions at both kitchen and customer/worker levels are introduced and discussed.

. Furthermore, considering the economic, environmental, and social costs associated with the phenomenon, the annual cost of global FLW is approximately 2.6 trillion dollars (FAO, 2014). Of the total amount of food produced, 13.8% is lost in the post-harvest, processing, and production phases, while 17%-equal to 931 million tons-is wasted at household, food service, and retail level (UNEP, 2021). Moreover, according to the latest Food Waste Index Report published by the UN Environment Program (UNEP), the average amount of household food waste (FW) is 74 kg per capita per year worldwide and similar values were observed between countries with low and middle income levels and those with high income levels, even if the distribution between the various stages of the production process is different. As regards the food service sector, it represents the third largest FW producer after households and food processing industries (FAO, 2019), with a global average of FW of 32 kg per capita per year (UNEP, 2021). At the European level, according to the estimates provided by the European FW project FUSIONS, this sector is responsible for 12% of total FW as it generates approximately 11 million tons of FW (Stenmarck et al., 2016).
In Italy, over 10% of FLW occurs during the agricultural and processing phases, approximately 20% during the distribution and sales phases while over 70% of FLW occurs during the consumption phases, that is, household and out-of-home FW, (CREA, 2020), which in total amounts to 15 billion Euros, four fifths of which (approximately 12 billion Euros) are thrown away within our homes, canteens, and restaurants.
Bearing the above issues in mind, out-of-home FW plays an important role considering that there is a clear trend toward out-of-home consumption, probably due to urbanization, rising incomes, conviviality, lack of time to prepare meals, and work commitments (d'Angelo et al., 2020;Eurostat, 2018;Lorenz et al., 2017;Pires et al., 2022;Principato et al., 2021b). It is important to understand the factors that lead to FW and carefully monitor the quantity of FW generated out-of-home. According to Hanson and Mitchel (2017), for every dollar invested in FLW reduction, a median site company realizes a $14 return, with food service companies obtaining the best performance.
Similarly, the same study found that companies that invest in FW reduction benefit both financially and non-financially. On the one hand, financial benefits mainly include a reduction in food procurement and disposal costs, a reduction in management and food storage costs, as well as a general improvement of the operational efficiency. On the other hand, non-financial benefits concern the enhancement and strengthening of the corporate brand and a favorable brand reputation as well as the growth of goodwill within the community (Hanson & Mitchel, 2017;Pearce & Berkenkamp, 2017).
To date, most of the studies on out-of-home FW have focused on restaurants, school canteens, hospitals, and social centers (Dias-Ferreira et al., 2015;Filimonau et al., 2020;Kowalewska et al., 2018;Liz et al., 2016;Oliveira et al., 2020;Papargyropoulou et al., 2016), while few studies have focused on company canteens (Jungbluth et al., 2016;Lassen et al., 2019;Lorenz et al., 2020Lorenz et al., , 2017Pires et al., 2022;Posch et al., 2022;Silvennoinen et al., 2019Silvennoinen et al., , 2015 by analyzing how better foodstuff management can be a win-win solution for company economic budgets as well as for environmental issues. This paper focuses on the out-of-home FW generated in a workplace canteen of an important Italian food production company, located in Pedrignano (Parma) in the Emilia-Romagna region, where many employees eat lunch every day. Aware of FW issues and impacts, the Italian food company launched a pilot project in March 2018, with the support of the digital Winnow platform to measure and minimize food wasted in its canteen.
The main aim of this paper is to assess whether and to what extent a workplace intervention, implemented with the aid of digital technology and aimed at reducing kitchen and employee FW, can contribute to defining best managerial practices and ways to achieve economic and environmental savings. More specifically, the operations implemented referred to accurate daily measurement and monitoring of FW and the identification of those areas with high FW levels to promote a greater level of sensitivity toward sustainability throughout the company.
The effectiveness and monitoring of the proposed actions are evaluated from a statistical and environmental perspective. First, a longitudinal analysis was conducted using machine learning tools such as Boosting, Random Forest (RF), and Bagging, which enabled us to identify the most important factors associated with the daily FW amount (measured both in absolute terms and relative terms considering the number of place settings) and to understand the areas in which measurement and monitoring interventions proved to be most effective or where more extensively interventions are required.
Second, life cycle assessment (LCA) enabled us to assess the environmental savings and how and to what extent a monitoring intervention using digital tools can contribute to sustainable business management.

Food waste at the food service level
The food service industry is responsible for producing and supplying meals to meet the nutritional needs of consumers, considering safety, sensory, and socio-cultural factors.
As previously mentioned, in recent years the number of meals served for out-of-home consumption has increased (FIPE, 2019;Pires et al., 2022) due to recreational reasons and the large number of people who eat in the workplace or student canteens. Contemporaneously, the amount of FW generated in this setting is growing exponentially, highlighting how this issue represents an important focus of research agenda (Principato et al., 2021b).
To address the topic of the current study, a systematic literature review of scientific papers included in SCOPUS database was conducted. In the first instance, the keywords combination used were "food waste" and "food service" and the timeframe was restricted to the years 2015−2022 (until October). Using the above-mentioned selection criteria 219 journal articles were extracted. Thereafter, to be more targeted with the theme of the study, the literature review was restricted to the keywords "food waste" and "canteen" or "workplace canteen." Again, the period was limited to the years 2015−2022 (until October), and in this case, 123 journal articles were identified. Each paper's identity elements (author(s), title, year of publication, journal, and abstract) were exported to an Excel file that was divided into columns. First, the papers obtained were grouped by year of publication, noting how the number of articles published yearly increased since 2015, probably due to rising awareness of the issue of FW ( Figure   S1 in Supporting Information S1). Second, since different types of food service were covered by the articles, they were classified by the type of food service they focused on. Here the set of articles has been reduced by removing those papers that were beyond the scope of the research, even if they were compliant with the above-mentioned selection criteria. The number of papers decreased from 123 to 102. Out of the articles identified, the majority have focused on school canteens (n = 69) and hospitality sector (n = 15) (which include bars, cafeterias, restaurants, hotels, and catering services), while only a few studies have centered on workplace canteens (n = 4). Moreover, n = 12 have focused on different types of food service, of which 4 have encompassed workplace canteens ( Figure S2 available in Supporting Information S1).
After evaluating the type of food service covered in each article, the set of articles was further restricted by focusing exclusively on journal articles that addressed the FW issue in relation to workplace canteens. Indeed, following this exclusion, the number of articles decreased from 102 to 8. Table 1 provides a summary of each paper, highlighting the type of intervention put forward in the study.
Based on the aforementioned literature review, the main causes of FW and its management in the food service sector, and especially in workplace canteens, will be highlighted within the following paragraph. Indeed, there are several causes for the generation of FW at this level and it can be explained by both service-related characteristics, such as service quality, and personal consumer variables such as food preferences. In this regard, according to the literature, the reasons why consumers waste food differ depending on their relationship with it, their culinary preferences, their emotional state, and their hunger during meals (Liz et al., 2021b;Lorenz et al., 2017;Principato et al., 2021b;Rohm et al., 2017). On the other hand, in relation to kitchen FW, the main reason to produce it is the difficulty to predict the amount of food that will be consumed each day, thus causing overpreparation and surplus food (Papargyropoulou et al., 2016;Principato et al., 2021b;Silvennoinen et al., 2019).
Focusing on FW at the company canteen level, a recent study carried out in a Portuguese workplace canteen identified excessive portions as the primary cause of employees' FW (Pires et al., 2022). This highlights the importance of constantly managing, planning, and monitoring the amount of food prepared and served to avoid wasteful overproduction which could be supported by measuring waste amounts (Silvennoinen et al., 2019). Moreover, a system of pre-ordering meal would allow a reduction of FW connected to meal planning and overproduction (Lorenz et al., 2020).
However, alongside better planning and monitoring, it is also important that consumers choose the appropriate quantities based on their appetite and requirements (Pires et al., 2022). In this sense, a possible measure to manage FW in worksite canteen could be to offer weight-based billing for the meal to ensure that employees become more attentive to their meal selections and more adaptable in deciding appropriate portion size (Lorenz et al., 2020). In this direction, Pires et al. (2022) found that consumers are sensitive to monetary incentives such as paying according to the quantity served and tend to be more careful and precise when deciding how much food to take on their plates.
Besides this, according to the literature, other factors that influence FW at workplace canteen are dislike of the taste, meal sensory characteristics, and menu acceptance (Pires et al., 2022). On the same line, Lorenz et al. (2017) determined that taste had a significant direct impact on plate leftovers of a catering company (Lorenz et al., 2017), stressing the relevance of improving food quality to influence FW from guests.
Finally, it is interesting to note that FW tends to increase when the meal offered is totally free (Ministero della Salute, 2018). In this regard, it is important to highlight that the meals offered in workplace canteens are not always complementary for the employees. For the sake of this study, it is important to note that the meal offered in the workplace canteen is partially paid by the company and partially by the employee for extras.
The conceptual framework used in this study to describe FW at food service level is in line with the framework developed by Principato et al. (2021) called the "Restaurant Food Waste Map" (RFWM), which aims at defining the stages in which FW occurs and at highlighting mitigation activities of FW generation processes in the restaurant sector. In fact, it identifies three phases in which the FW phenomenon occurs: (i) kitchen food preparation; (ii) food service, and (iii) client consumption (Principato et al., 2021a).
Kitchen FW occurs during the preparation phase mainly due to overpreparation, expired food, deterioration, cooking errors, peeling, cutting, etc.
Therefore, in this phase chefs and restaurant managers are those mainly responsible for the generation of FW.
Foodservice FW is defined as the food wasted during the service phase by staff as well as the food that has been prepared and has not been consumed. In this stage, all the staff (chefs, restaurant managers, and waiters) are responsible for generating FW.
Finally, client FW (CFW) concerns the scraps of food that clients leave on plates. In this case, those responsible for the waste are the diners themselves.
Therefore, our goal is to determine whether this framework that has already been used in the restaurant sector can prove useful in the food service sector, for investigating workplace canteens.

TA B L E 1
Set of papers focusing on food waste and leftovers in workplace canteens.

Digital tools to monitor food waste
Based on the above, understanding how much and why food is wasted by the food service sector is essential to implement measures aimed at reducing it. It is essential to measure the amount of FW generated, to identify the "problematic" issues of the process, to find new methods to reduce FW and verify their effectiveness. An adequate monitoring system would enable us to regularly measure the amount of waste produced both at the kitchen and the consumer level (Pires et al., 2022;Silvennoinen et al., 2019). Technologies are now key tools used to support FW reduction interventions (UNEP, 2021).
To enhance the understanding of how the use of digital tools could be used to support FW reduction intervention, a systematic literature review was conducted. We used SCOPUS database to identify relevant scientific papers using combinations of the following terms: "food waste" and "food service" and "digital." The period of analysis was focused on the years 2012−2022 (until October) by identifying 10 journal scientific articles. One of these articles was left out of our analysis since it did not mention any digital tools. Table S1 in Supporting Information S1 shows the main peculiarities of each study by highlighting the main aim, results obtained, context of application and type of digital tool used. It can be noted that seven of these articles concerned the use of digital photography as a digital tool to better manage FW.
According to the literature, digital solutions can aid FW prevention and minimization by monitoring and controlling food surplus and waste in the food service sector (Filimonau & De Coteau, 2019;Secondi et al., 2019;Strotmann et al., 2022). In this context, a recent study conducted on 91 German food service companies shows how digital tools enable us to reduce or prevent FW and how waste tracking tools can determine the quantity and origin of FW or the less popular dishes among customers, which can indirectly lead to process optimization, for example, by offering the most popular dishes or by reducing portion sizes (Strotmann et al., 2022). Moreover, Filimonau and De Coteau (2019 p. 242) stated that "smartphone apps have been developed to assist industry professionals in quantifying the volume and characterizing the content of food wasted with the design of subsequent mitigation measures." Moreover, the latest UNEP FW Index Report 2021 highlighted that efficient FW measurement and management systems have been developed such as Winnow AI tools, which use a camera and smart scales connected to data analysis and visualization systems to determine what foods are often thrown away and the data collected can be used by kitchen staff and managers to reduce FW.

The Winnow DT
The Winnow Waste Monitor is produced by a new technology company founded in 2013 in the United Kingdom that helps food service outlets track and reduce their FW by measuring and monitoring the wasted food thrown in the bins.
Thanks to the combination of artificial intelligence technology and a "smart" scale located under the FW bin connected to a tablet loaded with various menu items and food categories, it is possible to measure how much and which kind of food is wasted. The collected data provides staff with daily, weekly, and monthly reports which enables them to improve production processes, reduce FW, and cut costs by improving profitability.
With reference to corporate catering, the Winnow digital tool was used by numerous companies, such as IKEA, BASF, Réseau Santé Balcon du Jura (RSBJ), ESS-Compass Group, Wellcome Trust and Chartwells. In each of these realities, thanks to the use of Winnow, it has been possible to reduce waste, but up to now, there is not a scientific study on how this type of tool can help in the minimization of FW.
In our study, the implemented action concerned taking accurate daily measurements of the FW generated. More specifically, the project involved using a tablet for measuring and analyzing the FW and a five-stage implementation path as described in Figure 1.
Therefore, the weekly goal using this tool is to identify trends; carry out specific interventions to reduce kitchen, service, and employee leftovers; and set new incremental objectives to communicate with the team to involve them in the continuous improvement process.

The longitudinal data set: Definition and description
The F I G U R E 1 Food waste management routine to apply in workplace canteens.
F I G U R E 2 Absolute and relative (kilos over the meals served) food waste values over the weeks. The underlying data for this figure can be found in Supporting Information S2.
On a monthly basis, the highest (absolute) amounts of waste were generated in May and June 2018, as illustrated in Figure S2 reported in Supporting Information S1. The (overall) average monthly value of FW generated amounts to 670.06 kg for 2018 and 587.60 kg for 2019, as described in Figure 2.
The vertical red line in Figure 2 above divides the baseline from the rest of the data set. Both the absolute and relative values tend to decrease over time compared to the baseline. Another interesting aspect is that the absolute value of waste shows a decreasing trend even though the number of meals served generally shows an inverse trend.
Most of the waste (approximately 92% of the total waste) is generated by the uneaten food left on customers' plates followed by preparation waste while bread and expired food are the categories with the lowest percentage values. FW from plates represents 67.45% of the total volume of waste while preparation waste is 24.15%.
Both categories-FW from customers' dishes and preparation waste-show slight reductions over time, which is greater for preparation waste. value showed a greater reduction from 0.0424 to 0.0124, as illustrated in Figure S3 in Supporting Information S1. Finally, if we look at the reduction in economic terms of "Waste as a percentage of sales," it is important to note that this value was 2.9% in the baseline period, decreased to 1.2% in December 2018 while it still decreased in December 2019 to 0.6% of the total economic values of meals served (each meal has an approximate cost of 5 Euros).

Drivers and evolution over time: The machine learning (ML)-based analysis
In this section we introduce the three different ML decision tree models-Bagging, Random Forest (RF), and Boosting algorithms-used and for which the response variable was the overall volume of daily wasted food while the set of explanatory variables was: (i) the nine typologies of food wasted following the details presented in Section 3; (ii) the weekday; and (iii) number of place settings.

Bagging
Bagging, first introduced by Breiman (1996), is the short term to indicate Bootstrap AGGregatING and it represents a meta-algorithm (Ghojogh & Crowley, 2019) that can be used with any model both from a classifier and regression perspectives.
Let us assume that T indicates a learning (training) data set (y i , x i ), with i = 1, 2, . . . , N and a method used to construct a predictor f(x, T) using the given training set.
Bagging is based on a sequence of training set T k with (k = 1, . . . , K) and of the same size as T, generated by bootstrap (Tibshirani & Efron, 1993) selection from T.
Therefore, this general-purpose procedure aims at reducing the variance of a statistical learning model by using bootstrap re-sampling technique to obtain the sequence of training sets. Then, once obtained from each training set T k the estimatedf k (x), the aggregation of the estimates to obtain a single low-variance statistical learning models is obtained by averaging the resulting predictions from bootstrap as:

Random Forest
Focusing on RF methods, we used the Breiman's Random Forest algorithm (Breiman, 2001) which represents an extended specification compared to Bagging with the aim of de-correlating trees (subsets). As in Bagging, trees are grown with bootstrapping, but for each tree the model considers a random sample of M predictors, and not the whole set of predictors.

Boosting
Boosting is a method to transform a collection of weak classifiers into one strong classifier (Freund, 1995;Freund & Schapire, 1997;Schapire, 1990) that can be applied to statistical learning regression models and in our study, it was applied over decisional trees. This model does not build the tree through bootstrap, but instead comes as an estimate over a modified version of the original data. Indeed, a tree is estimated using the residuals r i of the previously estimated model as the independent variable instead of r i for all i in the training data set. After that, the new tree is added to the model to update the residuals. Indicating with r i the residuals and with b = 1, 2, . . . , B the number of trees, the model first fits the tree, then upgradê f(x) adding a reduced version of the tree:f then upgrade the residuals where is shrinkage parameters, which can further slowdown the learning process by allowing more trees to explain the residues. The final model will bef Trees can be very small, by adapting small trees the model improves slowly thus leading to define Boosting as a slow learning model.

Evaluating variable importance measures
For Bagging and Random Forest, we used two different measures, the Mean Decrease Gini (IncNodePurity) and the Mean Decrease Accuracy (%IncMSE), while for Boosting we referred to the Relative Influence measure.
Breiman (2001) proposed the first measure (IncNodePurity), using a weighted impurity measure to determine the importance of a variable to predict the independent variable, in this case by using the Gini impurity index.
The second measure (%IncMSE) is based on the decrease of mean square error (MSE) when an independent variable is removed from the model.
On the other hand, the relative importance used for Boosting is discussed by Hastie et al. (2009). During the tree construction, the split is done by minimizing the residual sum of squares (RSS), where each split causes a decrease of RSS. The relative importance of a predictor x p is obtained by considering all the splits due to this predictor in the sequence of trees and summing up the RSS percentage decrease observed due to these splits.

Assessing the environmental impact: The LCA methodology
The actions implemented at the food production company's canteens were also evaluated for their environmental loads using the LCA methodology.
According to the international standards of the ISO 14040 series (ISO, 2006a(ISO, , 2006b, LCA aims to analyze the environmental impacts and resource use in the life cycle of a product or service from raw material acquisition, production, use, and waste disposal. LCA is structured into four stages: (1) definition of goal and scope; (2) inventory analysis; (3) impacts assessment; (4) interpretation.
The overall goal of this study was to compare the environmental footprint associated with the actual FW production (baseline scenario) with those related to the years 2018 (2018 scenario) and 2019 (2019 scenario) when artificial intelligence tools were implemented in workplace canteens. "1 place setting/meal" was selected as the functional unit of all models. More specifically, the overall environmental impacts that occurred over the study period were determined and then a normalization by dividing each value by the number of meals served in that period was performed.
The system boundaries included the preparation of each foodstuff that may be wasted, the transportation of the catering waste, and its disposal in an aerobic composting plant.
Each foodstuff was modeled considering the cradle to consumer approach and including the following stages: farm or fishery, transportation to the processing industry, industrial processing, packaging, transportation from processing centers to distribution centers and from distribution centers to retailer, storage at retailer, transport from retailer to consumer, preparation at consumer, and disposal of packaging. Inventory data were collected from the recently opened Agribalyse 3.0 database for each food item or group. Agribalyse is the French life cycle inventory (LCI) database for the agriculture and food sector and the latest version, published in 2020, includes data for 2500 agricultural and food products. The data were not divided by food items for the "FW from plates" category, therefore the fraction of each food group (Table S1 in Supporting Information S1) was calculated from data reported in the study conducted by Vitale et al. (2018), in which data concerning the specific dishes consumed at lunchtime in a Barilla canteen in 2011 were recorded. The same approach was also applied to the "uncategorized food" category.
As regards the transportation step, the unit process for the transport was selected from the Ecoinvent database "Municipal waste collection service by 21 metric ton lorry {GLO}|market for| APOS, S," assuming a transport distance of 100 km. The composting process was modeled according to the process "Biowaste {CH}| treatment of biowaste, industrial composting| Cut-off, S," taken from the Ecoinvent database. Furthermore, a system expansion approach was followed to account for the valorization of the compost produced in all scenarios, which means that the environmental credits associated with its additional valorization were taken into consideration. It was assumed that compost replaces mineral fertilizers and its nutrient composition (1.4% of N, 0.7% of P 2 O 5 , 1.3% of K 2 O) was taken from Buratti et al. (2015). The compost yield on a dry weight basis, defined as the amount of compost produced from the total dry organic waste input, was assumed equal to 30% (Keng et al., 2020). The environmental impacts of mineral fertilizer production were taken from the Ecoinvent database (Table S3 in Supporting Information S1).
The LCA analysis was performed using the SimaPro software version 9.2.0.2 and the environmental impacts were determined by applying both IPCC 2013 GWP 100a (IPCC, 2014) and ReCiPe Endpoint v.1.1 (World ReCiPe H/A) (Huijbregts et al., 2017) methodologies. In particular, the ReCiPe Endpoint method enabled us to evaluate the damage to human health, ecosystem quality, and resources, and by following normalization and weighting steps, it is possible to calculate a single score result, measured in Ecopoints (Pt) to help decision makers and consumers gain a better understanding. Normalization and weighting are optional steps and could introduce additional uncertainties to the results of the LCA study.

F I G U R E 3
Bagging variable importance measures. The underlying data for this figure can be found in Supporting Information S2.
However, the application of these steps can facilitate decision making in situations where trade-offs among impact or damage category results exist. These steps are also crucial for providing support information to avoid subjective weighting of different impacts.

The ML estimation results: Bagging, RF, and Boosting
First, the data set was divided into two different subsets, a training set from April 2018 to September 2019 and a test set from October to December 2019. In this case, we did not consider the baseline, because the project was in its experimental phase at this period. We applied the model to consider kg of FW (daily) as the response variable and the day of the week, the number of meals served, and the type of waste as explanatory variables. The results were obtained by applying Breiman's Bagging model with the best number of trees to grow equal to 1000. Figure 3 shows that the most important drivers are represented by FW from employees' plates, preparation waste, and number of meals served.
The relative value of FW (kg/number of coverage) decreases over time, its absolute value (kg) shows a smaller reduction because it is positively related to the number of place settings, which increased over time, while the days of the week are not of significant importance. Figure 4 shows the variable importance resulting from Breiman's RF algorithm using an extension of Bagging. In fact, only a subset of all the independent variables in each tree was considered. In this case, the best number of trees was set to 500, while the number of variables to be considered in each tree that provides the best results is 10. The last two Figures (Figure 3 and Figure 4) appear to be very similar. Indeed, the most important variables are the same, while the MSE over the training data and the variance explained are higher in the RF.

The environmental impact
As described in Section 4.2, LCA was applied to determine the environmental impacts per "1 meal served" using two methods: IPCC 2013 GWP 100a, expressed in kg CO 2 -eq, which provided information about greenhouse gas (GHG) emissions, and ReCiPe Endpoint H/A World, expressed in points, to provide a comprehensive single score that measures 22 environmental impact categories, including acidification, ecotoxicity, mineral depletion, and land use.
As regards IPCC report on climate change impacts, a 100-year timeframe was assumed in this study. The results show that the implementation of artificial intelligence tools in workplace canteens led to a significant reduction in GHG emissions, from 18.3% in 2018 to 30.7% in 2019. Indeed, the F I G U R E 5 Boosting variable importance. The underlying data for this figure can be found in Supporting Information S2.  corresponding to a reduction in overall environmental impact equal to 32.7% compared to the baseline scenario. More specifically, human health was the principal cause of environmental degradation in all scenarios. Human activities are responsible for air pollution, such as particulate matter formation, ozone depletion, ionizing radiation, climate change, and photochemical oxidation. Approximately 90% of the total environmental impact was attributed to the human toxicity category in all scenarios, mainly due to wasted red meat and fish and mainly related to the GHG emissions from agricultural food production and particulate matter (PM) emissions from maritime transportation.
A sensitivity analysis was also performed to determine the effects of changing the scenario of catering waste disposal. According to ISPRA (2021), in Italy in 2020, 49.2% of the collected organic waste was disposed of in composting plants, 45.7% was treated by integrated anaerobic-aerobic methods and the remaining 5.1% was sent to anaerobic digestion. Thus, the sequential anaerobic-aerobic treatment of catering waste was assumed as an alternative disposal scenario. The anaerobic digestion treatment was modeled by using the process "Biowaste {RoW}| treatment of biowaste by anaerobic digestion | Cut-off, S" from the Ecoinvent database. A system expansion was applied to account for the environmental benefits of producing biogas (0.1 Nm 3 /kg of biowaste), assumed to be fed into a combined heat and power (CHP) plant. Electricity produced by the CHP was considered to replace the Italian grid electricity mix (Ecoinvent process: Electricity, high voltage {IT}| market for | Cut-off, S) while heat to displace the same amount produced by a natural gas boiler (Ecoinvent process: Heat, central or small-scale, natural gas {RER}| market group for | Cut-off, S).

DISCUSSION
The novelty of the present research lies in the fact that it analyzed and monitor the generation of FW in a workplace canteen using the innovative digital tool Winnow. Indeed, according to our literary review, concerning the area of study to date, most of the food service waste studies have been developed on school canteens and hospitality sector, while only few studies have focused on FW at workplace canteens (Lorenz et al., 2020(Lorenz et al., , 2017Pires et al., 2022;Posch et al., 2022;Silvennoinen et al., 2019Silvennoinen et al., , 2015. On the other hand, as regards the use of digital tools to monitor FW, the papers that address the issue of FW in the food service sector and that link it to digital innovations are very few (only nine) and, in most cases, the digital tool used is digital photography. Prescott et al. (2019), for example, utilize digital photography to evaluate the overall plate waste in a school's canteen. Similarly, Yoder et al. (2015) use pre-and post-meal digital photographs of students' school lunch to estimate fruit and vegetable waste. In addition to this, even in this case, very few studies have utilized digital methods to track FW in company canteens; instead, most have focused on school canteens, hospitals, and food services (Razalli et al., 2021;Secondi et al., 2019;Strotmann et al., 2022).
The results of this study show that the amount of FW in the workplace canteen decreased over time when the Winnow digital tool was used.
Preparation waste was reduced by 30% during the 2-year study period, while employees' food scraps were slightly reduced. These considerations suggest that it is necessary to raise employees' awareness through educational FW initiatives. Indeed, it has been demonstrated that by involving employees and developing educational programs and workshops to support managers and engage consumers in mitigation measures, FW can be effectively reduced (FIlimonau & De Coteau, 2019;Strotmann et al., 2017).
As regards the phases showing higher FW levels, the results reveal that most of the food wasted is due to employees' leftovers followed by preparation waste. Approximately 92% of the total FW occurs during these two phases, with FW from dishes representing 67.45% of the total FW volume, while preparation waste was 24.15%. As regards the food service stage, the lowest percentage values were observed for bread and expired food. These results are in line with those found by Principato et al. (2018), who analyzed FW in 127 restaurants located in Lazio and Tuscany (Italy) and found that the percentage of CFW was generally higher than that generated in the kitchens. Similarly, WRAP (2011) highlighted that the largest amounts of food are wasted during the food preparation and consumption stages in hospitality operations. However, these results are in contrast with those obtained by a Swedish study that found that the largest amount of food was wasted during the serving phase (Eriksson et al., 2017).
Another interesting result is the correlation between FW and the number of place settings for which a greater number of place settings generated greater quantities of FW. This result is in line with the empirical study conducted by Principato et al. (2018) according to which the restaurant size is a variable associated with FW and that the larger the restaurant the less CFW is generated. Contrastingly, the study published by Papargyroupoulou et al. (2016), conducted on a restaurant/hotel in Malaysia found that the volume of FW generated per customer decreased hand in hand with the number of customers served per day, probably due to economies of scale.
Finally, our study proves that FW reduction not only leads to economic savings for the company but also to lower environmental impacts which is in line with the triple-bottom-line approach of the Corporate Social Responsibility (CSR) of companies. In fact, according to our results, using DT to reduce FW in worksite canteens significantly reduced the GHG emissions, by 18.3% in 2018 and 30.7% in 2019 compared to the baseline scenario.
As indicated by Pearce and Berkenkam (2017), a company's commitment to FW reduction leads to the improvement in corporate water and GHG footprints caused by FW.

PRACTICAL IMPLICATIONS AND CONCLUDING REMARKS
The empirical research presented in this study stems from the will of a major Italian food company to monitor the amount of FW generated in their canteens. Although there has always been much debate about FW in the field of sustainable development and most studies have focused on household FW, the FW generated by individuals in company canteens has seldom been analyzed. The same cannot be said for the study of the waste generated in school canteens which has also been investigated using machine learning tools for programming (Malefors et al., 2021).
Overall, the results suggest that "informing actually means training." More specifically, it is essential to train employees and those involved in meal preparation, about the issue of daily waste to reduce it and raise people's awareness toward more sustainable consumption and management.
The usefulness of information-based interventions as a method to minimize FW has been observed in a variety of studies (Jagau & Vyrastekova, 2017;Manomaivibool et al., 2016;Reynolds et al., 2019;Schmidt, 2016;Young et al., 2017) including that of Reynolds et al. (2019), who found that information campaigns can reduce FW by up to 28%.
Although the educational and training tools are the most well known, other important interventions for reducing FW among consumers should be highlighted. Among these, changing the type of plate or portion sizes (Freedman & Brochado, 2010;Kallbekken & Saelen, 2013;Wansink & van Ittersum, 2013;Williamson et al., 2016), modifying the menu composition (Cohen et al., 2014;Martins et al., 2016;Schwartz et al., 2015), adopting technological solutions like smart fridges and cameras (Ganglbauer et al., 2013), or food sharing apps (Lazell, 2016;Lim et al., 2017), were all considered effective methods. On the same line, Dyen and Sirieix (2016) demonstrated how having cooking lessons helped customers manage their food and decrease waste. In addition to this, some simple interventions that may be implemented without significant investment or high maintenance could be color coding of shelving or sections in the refrigerator (Farr-Wharton et al., 2014), encouraging visitors at restaurants and large-scale homes to adapt their portions based on how hungry they are (Jagau & Vyrastekova, 2017) or to take lesser amounts at a buffet and return for more (Kallbekken & Saelen, 2013).
Future research should be aimed at analyzing the gradual re-entry of employees into the companies following the relaxation of COVID-19 restrictions and testing whether the new eating habits acquired at home during lockdown periods, together with the constant monitoring carried out, can really contribute to changing the behavior of individuals toward a more sustainable consumption (see Appendix S1 in Supporting Information S1).

ACKNOWLEDGMENTS
Open Access Funding provided by Universita degli Studi della Tuscia within the CRUI-CARE Agreement.