The approaches used in agro-food waste management were previously rudimentary at best, involving little more than the use of paper and pencil and even simple monitoring methods that could not cope with the modern supply chain challenges ( Table 1 ). Some of the existing treatments include composting, anaerobic digestion, and food redistribution but these techniques do not follow the current procedure of being efficient and effective ( Pereira et al., 2024 ). Modern research has revealed that somehow, incorporating technology into waste management can boost its performance. For example, organisations have recently installed more complex tracking systems that use IoT gadgets to track waste collection in real-time to enable timely action.

The use of IoT in managing agro-food waste has been considered in the past

few years. Smart sensors and RFID tags as some of the IoT devices are used to monitor different attributes of foods in order to determine qualities and wastage levels ( Sharma et al., 2022 ). Scholars suggest that these technologies can assist in minimization of spoilage by tracking and recording temperature and humidity in storage and transportation. In addition, IoT systems can enhance efficient tracking of stocks available and their time lifespan thus reducing unnecessary production.

They point out that big data analytics are fundamental for turning raw data from IoT devices into valuable information. Stakeholders can review a vast amount of data and find out the general tendencies concerning waste generation based on this analysis ( Rejeb et al., 2021 ). Literature review has revealed that big data can facilitate demand forecasting, the optimization of supply chain and the right utilization of resources. For instance, predictive analytics can be used to change the planting calendar with dates showing the history and markets to discourage overproduction ( Figure 1 ).

Advanced methods of artificial intelligence analysis are viewed as valuable approaches to reducing food losses in the agro-food chain. Several models have been created to categorize waste types; forecast future waste generation; and assess supply chain operational efficiency. Scholars underline the ability of ML models to predict the causes of food waste and give suggestions on how it can be prevented ( Munir et al., 2023 ). Also, machine learning reduces the work of finding appropriate data representation and depiction and helps to automate control decision-making in IoT systems ( Figure 2 ).

To reduce waste in the agro-food chain, many data sources are collected and used; satellite information for crop assessment, data from IoT devices, data which belongs to supply chains, and others. These features allow the achievement of the multiple goals set for the system, including the acquirement of data on different stages of the agro-food supply chain—from production to consumption ( Belaud et al., 2019 ). IoT sensors are used to monitor the environmental conditions in and around the devices and their operational parameters, which create conditions for waste generation. Mining is the process of extracting valuable resources from the Earth or analyzing large datasets to uncover patterns and insights. Mining helps the farmers to determine the appropriate time for harvesting by analyzing crop health and growth patterns on the farms through satellite imagery, which reveals regions of crop failure or diseases that cause waste ( Figure 3 ).

Information is collected from the logistics and distribution channels to determine where and when the products are moving, what stocks are available, and when they are set to expire so that wrong decisions on stock production are averted.

The data collected also goes through several pre-processing stages to make it useful for subsequent analysis. These steps include

Cleaning—This is because data accuracy is important in any dataset, and so cleaning it to reduce improper entries is important. This may be done to rectify wrong records, or AVs that may have been omitted or complied with in different formats.

Filtering—It is always helpful to choose what pieces of information contribute to waste analysis to highlight the most influential factors. For example, one can remove noise, or non-relevant data can help improve the meaning of the insights that can be gathered from this data.

Feature Extraction—Important factors that define waste generation should be clearly described and defined for the analysis ( Arshadi et al., 2016 ). It may involve obtaining characteristics like the fluctuations of the market demand, the seasonal characteristics, or such environmental parameters as related to the high throughput of wastes.

Using processed data, several analytical techniques are used for the analysis of waste trend determination. Statistical methods including regression analysis, clustering, and time series are used to analyze the data and make forecasts on future waste generation. The extension of machine learning algorithms in the end also increases the efficiency of finding varied patterns in the data.

Descriptive for the interactions between different variables such as temperature with spoiling rates, regression is useful in scaling the impact of one factor on waste ( Oruganti et al., 2023 ). This technique simply clusters observations with the aim of finding patterns of waste generation in other related contexts, such as the type of product or geographical region. Clustering can then be used to identify what products are likely to suffer from waste in the given circumstances. This method of waste generation prediction involves the use of historical data over time, in which the foreseen waste generation is analyzed with reference to the existing season and market forces. It enables the stakeholders involved to make appropriate changes to the production calendar actively ( Table 2 ).

A food processing plant that was especially in fruit and vegetable processing required a big data analytics solution to enhance its production calendar and minimize wastage. Before the adoption of this system, this plant used to experience cases of spoilage since there was wrong stocking, and since the demand for some foods fluctuated, the plant used to experience a lot of wastage.

Background—To add, the plant employed in the processing of fruits and vegetables is more than fifty tons per day, of which approximately fifteen percent are chucked as waste because of spoilage. This translated to about 7.5 tons of waste daily, which in a year translated to over $1 million in losses ( Koroteev et al., 2022 ).

Implementation—They had IoT sensors put up across the stored and processed commodities to allow for continuous tracking of environmental conditions. Information received from these sensors was physically stored in a cloud based big data storage structure where machine learning based models were implemented for predictive analysis.

Results—The control plant (specific facility or operational unit within the organization where the cold chain system was implemented and monitored to assess its impact on spoilage rates, cost savings, and workflow efficiency) also recorded a decline in spoilage rate in the first six months from 7.5 tons per day to 5.25 tons because of the application of the cold chain system. This reduction not only proved to be an annual cost saving of an estimated $300,000 but also proved quite effective for the overall workflow efficiency of the company as it provided much required flexibility to perfectly match the production strategies to the market targets.

Key Metrics: Waste Reduction—From 15% to 10.5% of total output; Cost Savings—Estimated $300,000 per year; Operational Efficiency—Improved production scheduling and reduced labor costs by approximately 10%.

One very large grocery store chain put into action a big data plenty predictive analytics and system to better control inventory and decrease spoilage in its outlets.

Background—Although the chain had been managing over 1000 stores all across the United States, the organization found its average food waste percentage of 8% across its perishable food categories, resulting in a yearly loss of about $600 million.

Implementation—Machine learning models were created based on historical sales data in conjunction with real-time data from IoT-enabled smart shelves to make better predictions of demand patterns by the retail chain. Some of these models included external factors such as changes in seasons, social events and even the weather ( Sridhar et al., 2021 ).

Results—For perishable foods sub-sector, reduced food wastage by a quarter within the first year of applying the theory to the retail chain. It translated to about $150 million in cost savings per year. Also, the gains made in the inventory through purchases inventory management caused stock replenishment activities to minimize stockouts by 20 percent.

Key Metrics—Waste Reduction—From 8% to 6%; Cost Savings—Approximately $150 million annually; Stockout Reduction—Improved from a rate of 15% to 12%.

A case of a large-scale farm implementing the farm management IoT Big Data Solution to reduce water wastage and draught on crops through right irrigation.

Background—The farm was overseeing more than two thousand acres of crop production that were recording high losses because of poor water management and control; fresh crop failures were costing the farm about $200,000 per year.

Implementation—In this case, the farm had placed soil moisture sensors that relayed the condition of the soils to the farm at different fields. This data was used along with the seven-day forecasts on the weather patterns to come up with machine learning algorithms that helped in the prediction of the irrigation patterns based on the moisture levels of the soil ( Tummers et al., 2021 ).

Results—When this system was established on the farm for one growing season, this farm claimed that they experienced a decrease in crop damage by about 40% due to water content. From this improvement, they estimated that their yield would be boosted by approximately $80,000 per year.

Key Metrics—Crop Loss Reduction—$200,000 annually, which is reduced to $120,000 annually; Yield Increase—Potential increase in annual revenue of about $80,000.

The conclusions drawn from the case studies highlight how big data analytics and machine learning can create real added value in agro-food logistic systems and processes. Through real-time data and data-driven prediction, organizations are able to shave off wastage while improving operations, profitability and effectiveness. ( Table 3 )

There are several machine learning programs that address specific issues linked to agro-food waste. They are models which are used to classify types of waste for

instance depending on its source, content or spoilage factors ( Kumar & Singh, 2024 ). However, other forms of machine learning such as classification methods such as decision trees and support vector machines can predict which products are susceptible to being wasted given the past record and conditions of the environment ( Bhuiyan et al., 2023 ). Regression Models are used to estimate the amounts of waste produced under conditions such as changing temperatures and humidity in the storage environment. Methods such as linear regression as well as significantly more advanced models like random forest can express the regularity of the input variables to the waste output which can help to collect the necessary inventory information and make decisions among the stakeholders ( Rahaman et al., 2023 ; Islam et al., 2024 ). Moreover, clustering analysis is used to help categorize similar waste patterns so that relevant factors causing waste can be determined. Methods such as the k-means clustering or hierarchical clustering provide a way to identify which product or process is most susceptible to waste with subsequent steps to organize an intervention ( Bhuiyan et al., 2023 ).

The idea behind Ensemble Methods is to train multiple models so that when used together, we get better predictions with better accuracy and stability. Also, somewhat related to mixing the values received from different models, bagging and boosting can improve the performance of individual models and can offer a more inclusive outlook into waste dynamics ( Lakhouit et al., 2023 ).

Hence, each of the models is constructed with a special architecture to deal with the problem under consideration. Closely related to the deep learning models, it is necessary to determine the number of layers and neurons. A well-designed neuron circuit diagram might contain a number of hidden neurons that help to recognize the high-level dependencies in the data while the neurons in each sequential layer capture more detailed dependencies. Metrics about individual learning rates, participation rates, and regularization during training contain sensitive information about such learning models ( Liu et al., 2017 ). To choose model parameters we can use a grid search or random search which can help across hyperparameters in order to get the best performance of a given data set. Feature engineering is the method of defining, changing, or developing new attributes required by a model to be optimized. Feature engineering is demonstrative of the procedures that can mean improving the potential of a model to forecast waste, especially using domain knowledge of the agro-food processes. Techniques like k-fold cross validation also make it possible to guarantee that the algorithm performs well on new datasets. Overfitting is prevented through this approach because the performance of the model is checked through another portion of the training data ( Rajvanshi et al., 2023 ).

The training process involves

Using three sets of data—training, validation, and a test set—is crucial to properly judging the prowess of a given model. The training set is then used to build the model, and the validation set is used to tune the parameters of the model, and the test set is used to test the model. In the case of cross-validation, the idea is to get some level of confidence in the model by assessing it on different subsets of data. This makes it easy to detect early signs of over and under-fitting during the training process, and it can be effectively used by designers. To identify effective models, measures like accuracy, precision-recall, F1-score, and AUC will be useful in providing a complete account of the performance of the models ( Golbraikh et al., 2003 ). These have the effects of enabling the stakeholders to evaluate how accurate the constructed model is in predicting waste and the ability to reduce waste in the future. Some tools that help understand the relative importance of certain features when it comes to the output are SHAP (SHapley Additive exPlanations) values or LIME (Local Interpretable Model-Agnostic Explanations).

The results section will discuss the comparison between the various models in the aspect of how well and how effectively they can estimate and reduce different forms of agro-food waste. It will also make possible conclusions about one model or the other that was most efficient under certain circumstances. Specific differences between multiple machine learning models will be discussed by specifying their accuracy, speed, and applicability for various contexts in the agro-food sector ( Montecinos et al., 2018 ). A summary of the strengths and weaknesses of various approaches will be discussed in this work. The findings will therefore present two aspects, these are: which of the features was more relevant for forecasting waste generation? For instance, some environmental conditions or operational practices may develop into influential factors determining the rates of waste. They shall also be jointly accompanied by implications in relation to organizational stakeholders in the agro-food supply network, including guidelines on how dwelling and deploying machine learning solutions they are likely to encounter are best affected. This portion of the paper broadens the presentation of machine learning systems for reducing agro-food waste and highlights the significance of building solid structures in addition to considering the practical application of already developed models in the current setting ( Aït-Kaddour et al., 2024 ).

Novel developments in both the field of BDAS and machine learning have created new possibilities for improving agro-food waste management. Improvements like remote control of monitoring systems by IoT have also evolved to a level where changes can be managed in real time. New technologies enabled by tiny, sophisticated sensors for temperature and humidity, as well as gas emission, spoilage, and many other parameters, changed the approach to waste management ( El-Ramady et al., 2022 ). These sensors produce detailed information that can help to develop intervention schedules. Cloud computing, when integrated with IoT and BDAS provides the means of storing and processing a large quantum of data that arises from the agro-food processes. This also promotes solutions that can easily be scaled up and be accessed anywhere and standards ensure that the stakeholders can make proper decisions upon getting real time information. Featuring the ability to provide indisputable trails of merchandizer’s movement, blockchain can increase competence in waste control in the food chain. This technology can show areas that need improvement so that proper changes can be made. Competence in the manufacturing of AI algorithms to analyze big data sets and decision support systems is gradually being used in different fields ( Procentese et al., 2019 ). Such developments provide the possibility to control the equipment’s condition without its failure, efficient deliveries, as well as better demand estimation leading to minimized wastage.

In this research, new model configurations are proposed to utilize both real time IoT data streams and past data. They have become useful tools to improve prediction and organizational effectiveness within waste management systems. Hybrid models can consider more patterns in generation data in comparison with the best regression and classification models, since the last ones combine different types of machine learning algorithms, for instance, regression and classification models integrated into one ensemble model ( Kotsiantis et al., 2006 ). This approach improves the reliability of the predictions. It is a strategic necessity to integrate models that can be adjusted to respond to changing conditions in real-time into the waste management sphere. These models rely on real-time data input from IoT technologies to refine estimation in real-time operational conditions, resulting in enhanced waste estimation. IoT systems need to be integrated with machine learning solutions that give a broader picture of the agro-food supply chain ( Das, 2019 ). This integrative approach helps in making good decisions since more than one factor influencing generation of waste is considered.

IoT is used to collect data that serves as the foundation for machine learning by feeding real-time data. The following section will also describe how it is possible to catch real-time data feeds from sensors and fit them into analytical structures that make possible dynamic decision-making. Applying data fusion procedures makes it possible to include data from various sources (for example, IoT, weather conditions, customers’ demand) into a single analytical system ( Joshi et al., 2022 ). This rich set of data improves the reliability of machine learning forecasts of waste production. It also allows stakeholders to address new issues when they are identified by analyzing data from the stream as it happens. For example, when there is a high temperature within storage facilities as obtained from the sensors, the appropriate corrective measures must be taken to avoid spoilt food. The integration of feedback on IoT systems and machine learning models guarantees continual optimization and criminal intentions of individuals they will detect, and as new data is collected, the model is trained further to improve its accuracy.

An exploration of how the adoption of IoT and machine learning could lead to a future where the actual reduction of waste could be at least partly automated will be made. Automate the process so that changes can easily occur in real time and lead to better resource allocation as supported by the predictive analytic tools. When machine learning is applied to IoT data sources, organizations can create decision-making systems that are capable of processing data in real-time when it comes to things such as inventory control, production planning or logistics ( Aziz et al., 2023 ; Rahaman et al., 2023 ). For instance, the automation of a specific process could entail changing order quantities depending on a predicted, necessary demand rate, calculated from previous sales trends. Automation is not limited only to changes in operations; it includes predictions for maintenance and preservation of food machinery as well. In observing performance data of equipment, an organization is likely to identify some failures before they happen hence reducing on time wasted. Introducing bins that incorporate sensors that determine their fill levels will make collection time for waste disposal services efficient ( Engler, 2004 ). This eradicates the time-wasting as well as the cost-incensing exercise of making many trips to the compound to remove the waste just in case it has overflowed. These advancements are important innovations to properly address the current wastage practices, mostly within the agro-food sector, to demonstrate that technology can promote improvements, promote sustainability efforts, and promote a reduction of wastage within supply chain networks.