Predicting Postharvest Food Losses at National and Sub-National Levels Using Data-Driven and Knowledge-Based Neural Networks

Abstract

Food loss is a major challenge for global food security, resource use efficiency, and sustainability. However, collecting primary food loss data is costly. This study explores a neural network-based approach to estimate food loss in the postharvest stage using the FAO’s food balance sheets for proof of concept. We investigated both traditional data-driven feedforward neural networks (FNNs) and knowledge-informed neural networks (KiNNs) using rice, wheat, and apple data from the FAO’s food balance sheets. The results show relatively high prediction accuracy with the data-driven approach when a larger amount of data is available. It also demonstrates the high potential of using KiNNs to improve the prediction accuracy when data availability is relatively limited. In general, the proposed approach shows great potential to be developed into an effective supplementary tool that can partially replace costly primary food loss data collection at the postharvest stage, which is particularly valuable when resources for primary data collection are limited.

1. Introduction

Food loss and waste (FLW) has emerged as a critical issue, which draws substantial attention from policymakers, researchers, and industry professionals all over the world. According to the Food and Agriculture Organization (FAO), between 25% and 33% of food produced for human consumption is either lost or wasted globally. This volume is sufficient to feed approximately 2 billion people. Given the projection that the global population will reach 9 billion by 2050 [1], addressing FLW is vital to bridging the gap between food supply and demand and thereby enhances global food security [2,3]. Beyond its social and ethical implications, FLW also brings in significant economic and environmental challenges. The FAO estimated in 2013 that the global economic value of FLW was approximately USD 750 billion in 2007. FLW has become a significant environmental issue in terms of resource use efficiency (e.g., water and land use), climate change, biodiversity losses, etc. For example, Porter et al. (2016) [4] calculated that greenhouse gas (GHG) emissions associated with FLW during the production phase in 2011 amounted to 2.2 gigatonnes (Gt).

To address this urgent issue, the United Nations included FLW reduction as part of its Sustainable Development Goals (SDGs) (https://www.undp.org/sustainable-development-goals, accessed on 6 January 2025), aiming to halve per capita FLW at the consumer and retail levels by 2030 (e.g., [5]). Achieving this target requires the implementation of carefully designed FLW intervention strategies with robust evaluation frameworks to measure their impacts. However, the development of effective evaluation mechanisms remains in the preliminary stage [5], largely due to insufficient data for tracking progress in FLW reduction over time [6]. Even though the Food Loss and Waste indexes were developed by the FAO and UN environment programme, FLW data are still limited at the national and sub-national level due to the high costs associated with primary data collection. This lack of reliable data makes it significantly challenging to evaluate the effectiveness of intervention strategies.

The existing literature for FLW quantification at national and sub-national levels primarily focuses on the developed countries, particularly in the USA and Europe (even though EU countries do collect and share the FLW data on a national level per chain link, there are limited studies for cross-validation), with an emphasis on the retail and household stages (e.g., [7,8,9,10,11,12,13,14]). This means that they essentially look into the “food wastes” (FWs) instead of the “food losses” (FLs), which are distinct concepts depending on whether they belong to upstream or downstream segments of the food chains (e.g., [15,16]). Compared to studies on food waste (FW), research addressing FL is relatively limited. Examples include the farm-level loss studies (e.g., [17,18,19]), with even fewer focusing on the postharvest stage (e.g., [4,20,21,22]). This may be due to the fact that, unlike food waste, which is often a hot topic in higher-income countries with richer resources for primary data collection, food loss is predominantly an issue in lower-income countries, where resources for such data collection are significantly more limited.

Although primary data collection for FLW is costly, it is surprising that there are still very few studies leveraging AI and machine learning to address data gaps at the national and sub-national levels, despite the current quick development in AI advancements. A recent attempt by Nijloveanu, Tița [23] focuses on FW prediction in Romania. The authors propose using a Decision Tree Approach (DTA) and a Neural Network Approach (NNA) to predict outcomes for specific FW agents (e.g., consumers) based on pattern recognition from the collected data. In the field of quality-decay modelling (especially fruit and vegetables), which predicts product deterioration and occasionally may be able to be used to calculate FLW as a byproduct (e.g., expired amount), AI models are more frequently utilized (e.g., [24,25,26,27,28]). However, these studies often focus on specific supply chains with very limited datasets that are typically not very representative. Additionally, the data to train the model are often collected under experimental conditions, which may make the model not accurately reflect real-world situations.

This paper aims to fill the knowledge gap in the stream of the literature where AI and machine learning approaches are employed to predict the postharvest FL at the national and sub-national levels, combining publicly available secondary data sources to supplement or potentially even replace the costly primary data collection to a large extent. We explored both traditional FNNs and KiNNs as machine learning models to link the predictor variables with the target variable. FNNs, also known as multilayer perceptrons (MLPs), are a fundamental type of neural network consisting of multiple fully connected layers that connect the input layer to the output layer [29]. This data-driven approach is well known for its ability to capture complex nonlinear relationships within the data. Unlike purely data-driven FNNs, KiNNs are a form of knowledge-based AI that embeds domain knowledge into the model [30]. This approach constrains the neural network’s parameters using knowledge-based rules, making it particularly useful for extrapolation tasks when data availability or diversity is limited.

To the best of our knowledge, we are the first to use the publicly available secondary data to do this kind of modelling to fill in the postharvest FL data gap. Moreover, we are the first in this field to demonstrate how domain knowledge can be incorporated into neural networks to improve prediction accuracy through KiNNs, which is especially useful when data availability decreases. Furthermore, it is important to note that the primary contribution of this paper remains at the conceptual level. Our aim is to demonstrate the added value of the proposed approach by leveraging FAO food balance sheets and domain knowledge to address data gaps in postharvest FL estimation. As such, this study presents a prototype model with a proof-of-concept analysis, and no effort was made in fine tuning the models or identifying optimal hyperparameter settings. Finally, our research also contributes to the advancement of the circular economy. Accurately predicting postharvest losses provides a solid foundation for reusing these materials in circular applications, such as feed production. Moreover, this work opens up new research directions for similar tasks in the field of FLW monitoring and valorization, further supporting circular economy objectives.

The remainder of this paper is organized as follows: Section 2 presents the use of the traditional data-driven FNN to predict postharvest food losses using a larger dataset. Section 3 introduces the knowledge-based KiNN approach to address the challenge of data scarcity by incorporating domain knowledge into the neural network. Section 4 gives a general discussion of the research findings.

2. Data and Basic Feedforward Neural Network Model

As aforementioned, the availability of national and sub-national FL data at the postharvest stage is very limited in the existing literature. However, the FAO’s food balance sheets (FBSs) provide a highly valuable training data source at the country level with labelled target variables, which can be leveraged to bridge this knowledge gap at the national or even the more granular subnational levels. This is feasible because the dataset includes a wide range of countries with diverse sizes, economic conditions, and food system characteristics, many of which can reasonably be compared to the variations seen across different sub-national regions within a single country. In this context, country names serve as more nominal labels rather than different levels of categorical variables, which have essential modelling meanings. As long as the predictor variables used in the model are relevant at both national and subnational levels, it is viable to train the model on the national-level data and apply it for the national and even sub-national inference, even if direct sub-national data are unavailable. This methodology enables more precise, localized insights into the postharvest food loss patterns, which can help to fill critical data gaps in food security research and policy design.

Specifically, we used the values of “Losses” from the FAO food balance sheets, which represent postharvest FL, as the ground-truth target variable for training and testing the AI models. For each product group and year, we divided the FL by the population of the country, to derive the annual per capita FL as the target variable in the FNN model.

Figure 1 shows a simplified architecture of the FNN model with the predictor variables, which include an input layer, hidden layers, and an output layer. In this purely data-driven model, the only loss minimized by the neural network is the deviation between the predicted value and the ground-truth value of the target variable, which is different from the KiNN model (as presented in Section 3). The loss function used in the FNN model follows the standard Root Mean Squared Error (RMSE) formulation, as shown below:

$${RMSE}_y=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$$

where

• n is the number of total training datapoints;
• $y_i$ is the ground-truth value of the target variable for the ith training datapoint;
• ${\widehat{y}}_i$ is the predicted value of the target variable.

During the backpropagation process, the loss function guides the optimization of the FNN’s parameters with a single objective: minimizing ${RMSE}_y$. ${RMSE}_y$, also referred to as “data loss”, reflects the model’s performance based solely on the labelled target variable y using ground-truth data. In contrast, the KiNN model incorporates both “data loss” and an additional “rule loss”, which enables it to learn not only from data but also from embedded domain knowledge.

The list of predictors used in the input layers is presented in

Table 1. The predictors of the model and their explanations.

Predictor	Explanation	Source
FAO Region	The seven regions (defined by the FAO according to the geography and development phase) to which a nation or sub-national regions belong. For example, India is a country belonging to “South and South-East Asia”	FAO Food Balance Sheets
Year	The year that the predictor and target variables refer to	N/A
Population (1000 persons)	The number of the population of a nation or a sub-national region. For example, India had a population of 1,396,387.13 people in 2020	FAO Food Balance Sheets
GDP_pc (USD)	The GDP per capita of a nation or a sub-national region. For example, India had a GDP per capita of USD 1915.551588 in 2020	World Bank: DataBank
Precipitation (mm)	The precipitation (mm) of a nation or a sub-national region. For example, India had a precipitation value of 1083 mm in 2020	World Bank: DataBank
Protein_supply (g/capita/day)	The per capita protein supply of a nation or a sub-national region. For example, India had a protein supply of rice measuring 13.62 g/capita/day in 2020	FAO Food Balance Sheets
Production (1000 t)	The production of the referred crop of a nation or a sub-national region. For example, India produced 186,500 tons of rice in 2020	FAO Food Balance Sheets

with their definitions and sources. The categorical variable “FAO Region” is transformed into the numerical value using the standard one-hot encoding. The predictor files are then merged with the target variable file using an inner join to create the dataset for modelling.

The reasons for us to select those predictor variables are based on data availability and their potential (direct or indirect) relevance for FL prediction. It is important to note that this research serves as a proof of concept to demonstrate the potential of the proposed approach; the authors do not intend to exhaustively identify all relevant predictors or develop a fully optimized model at this stage.

For cross-validation, the whole dataset is split into a training and testing dataset to avoid data leakage. Since our target variable is continuous, we evaluated model performance using the coefficient of determination (R²) between the predicted and ground-truth FL values on the test dataset, rather than relying on mean squared error (MSE) or RMSE. This choice was made because R² reflects the proportion of variance explained by the model and is less dependent on the absolute scale of the target variable, unlike the MSE and RMSE.

To further rule out the random effect of one-off arbitrary dataset splitting, we randomly split the original dataset 50 times. In this way, instead of obtaining one R² value, we derived a distribution of 50 R² values (see Figure 2).

For the proof of concept, we selected rice, wheat, and apple to test the feasibility of the proposed approach. Among these, rice was chosen for a more detailed analysis, which includes the ablation study to evaluate the influence of individual predictors on the model’s accuracy by systematically removing one predictor at a time from the full model (i.e., the model including all predictors). This approach allows for an objective assessment of each predictor’s impact by comparing the performance of the reduced models with that of the original.

Table 2. The original and reduced model naming with the explanations for rice.

Model Name	Explanation
All_predictors	The original model with all predictors described in Table 1
FAO_Region_remove	The reduced model with “FAO Region” removed
Year_remove	The reduced model with “Year” removed
Population (1000 person)_reomve	The reduced model with “Population (1000 person)” removed
GDP_pc_remove	The reduced model with “GDP_pc” removed
Precipitation (mm)_remove	The reduced model with “Precipitation (mm)” removed
Protein_supply (g/capita/day)_remove	The reduced model with “Protein_supply (g/capita/day)” removed
Production (1000 t)_remove	The reduced model with “Production (1000 t)” removed

lists all the models involved in this ablation study for rice. Focusing on rice, with lighter analyses on wheat and apple, helps limit computational time while remaining adequate for demonstration purposes. To further balance computational efficiency with sufficient model learning, all models listed in Table 2 were trained for 1000 epochs.

Finally, the dataset used in this section covers the period from 1961 to 2020 with yearly data, representing a relatively large dataset compared to the one used in the KiNN section.

Figure 3 presents the performance distributions of R² for all models listed in Table 2, which gives a detailed overview of how each removed predictor influences the model’s prediction accuracy on the test dataset. The original FNN model with all predictors achieves a high average R² value close to 95%. This demonstrates its strong capability to capture underlying data patterns through the combined predictive power of the input predictors. This high R² value shows that the model effectively explains the variance in the data and suggests the successful learning of complex relationships between variables.

The ablation study further demonstrates the contribution of each predictor by evaluating model performance with an individual predictor excluded each time. Prominently, predictors such as “Year”, “Precipitation (mm)”, “GDP per capita (USD)”, and “Production (1000 t)” contribute significantly to model accuracy. When these variables are omitted, the model’s performance drops clearly, which indicates that they capture essential patterns in the data related to time trends, weather conditions, economic situation, and production volume. This suggests that these predictors have meaningful relationships with the target variable and play critical roles in enabling the model to predict accurately. Moreover, removing “GDP per capita (USD)” and “Precipitation (mm)” also increases the variance of the R² distribution. This indicates that these predictors not only affect the model’s average accuracy but also play a crucial role in reinforcing the underlying relationships between the predictor and target variables learned by the model.

On the other hand, predictors like “FAO Region” and “Population (1000 persons)” show minimal impact on the model’s performance when excluded, implying that these predictors do not add much additional information to the model. This lack of effect could indicate either that these variables do not have a significant relationship with the target variable or that other predictors already capture the relevant information these variables can provide and adding them only introduces more noise to the model. Interestingly, ‘Protein Supply (g/capita/day)’ has a negative impact on model performance. This may be due to the fact that protein supply from rice includes not only domestically produced rice, which is directly linked to postharvest losses, but also imported rice, which is not directly linked. This discrepancy may introduce noise into the model and reduce its predictive accuracy. Therefore, ‘Protein Supply (g/capita/day)’ may not be a suitable predictor for the model and should be excluded when applying the model in practice.

Figure 4 presents the R² distributions for models incorporating all indicators for rice, wheat, and apple. The visually different shape of the rice distribution compared to that in Figure 3 is due to differences in axis scaling. The average accuracy of the wheat model is lower than the rice model, and the apple distribution has the lowest average accuracy. The variance of the distribution is inversely related to the average accuracy. This may suggest that the low predictability may stem from high variability specific to the product or postharvest system for wheat and apple, compared to rice.

3. Knowledge-Informed Neural Network Model

The FNN demonstrates relatively good performance when the dataset is sufficiently rich compared to the variance that it needs to explain. However, in situations where data are limited relative to its variability, a knowledge-based AI approach becomes necessary to incorporate domain knowledge to enhance the learning efficiency of the neural network. This strategy aligns with the principles of scientific machine learning (https://sciml.wur.nl/reviews/sciml/sciml.html, accessed on 8 February 2025), which is a domain originating from the physics field with the so-called physics-informed neural networks (PiNNs). This type of AI model uses the domain knowledge to guide the learning process of neural networks (e.g., [31,32]) and have been shown to be more efficient than traditional, purely data-driven neural networks like FNNs. Although originally developed in the field of physics, the application of PINNs has extended far beyond physics and is now being explored in many other domains. For example, Bae, Kang [33] applied a PiNN in the field of finance to predict option prices and local volatility surfaces. Naderi, Perez-Raya [34] used a PiNN to predict the spatio-temporal chemical concentration in the environment and claim it is a significant advancement in environmental monitoring technologies. Okawa and Iwata [35] claimed that they transported the concepts of PINNs from natural science (i.e., physics) into social science (i.e., sociology and social psychology) to predict opinion dynamics via the so-called sociologically informed neural networks.

Different from the earlier ad hoc studies, Faroughi, Pawar [36] provided a methodological generalization in the field of physics by proposing a typology that distinguishes between physics-guided, physics-encoded, and physics-informed neural networks within the context of scientific computing. The authors point out that a common approach to implementing physics-informed neural networks is by incorporating a physics-based penalty term into the loss function. Building on this, Guo, Harbers [30] further refined and broadened the typology by extending it beyond physics to encompass general domain knowledge. They introduced the term KiNNs to generalize this class of models. The KiNN model developed in this study follows the definition proposed by Guo, Harbers [30], in which a knowledge-based “rule loss” is incorporated to guide the neural network’s learning process. In this research, we build on the aforementioned literature by developing a KiNN model and comparing it with a corresponding FNN model to intuitively demonstrate the added value of incorporating domain knowledge.

It is well known that fruits require cold storage to maintain their postharvest quality, which makes the level of cold chain infrastructure a key factor influencing postharvest losses. Based on this established knowledge, there is a significant difference in postharvest fruit losses between developed (high-income) and developing (low-income) countries. This insight can be leveraged to guide the model’s learning process through the incorporation of knowledge-based constraints. For this sake, we selected the apple dataset for the models developed in this section. However, to demonstrate the added value of the KiNN compared to the traditional data-driven FNN, we used only a subset of the apple data, covering the period from 2001 to 2020, to train and test the two models developed in this section: a KiNN model and a corresponding FNN model. Figure 5 shows the simplified schematic representation of the KiNN.

Different from the FNN model (Figure 1), an additional output head z is introduced to account for the knowledge-based rule. z is a binary variable with 1 indicating the higher-income countries and 0 indicating the lower-income country. To ensure consistent country categorization, we used a GDP per capita threshold (since our research serves as a proof of concept to demonstrate the added value of domain knowledge, we did not treat this threshold as a hyperparameter to optimize the model performance) of USD 20,000 (in 2020) to distinguish between high-income (z = 1) and low-income (z = 0) countries. The loss function of the KiNN model is shown as follows:

$${Loss={RMSE}_y+RMSE}_z$$

$${RMSE}_z=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(z_i-{\widehat{z}}_i\right)}^2}$$

where

• n is the number of total training datapoints.
• $z_i$ is the ground-truth value of the target variable for the ith training datapoint.
• ${\widehat{z}}_i$ is the predicted value of the target variable.

The loss function now contains both a “data loss” term ${RMSE}_y$ and a “rule loss” term ${RMSE}_z$, the latter of which incorporates prior knowledge that high- and low-income countries exhibit significant differences in postharvest losses for fruits and vegetables. In this case, the neural network is expected not only to accurately predict the primary target variable y but also to effectively predict the additional output z. In other words, the network should be able to distinguish between high-income and low-income countries within the context of food loss prediction. This essentially imposes an additional constraint on the neural network, reducing the degrees of freedom in its parameters compared to without the second output head. As a result, the network must carefully balance minimizing the data-driven loss with adhering to the knowledge-based rule loss, ultimately aiming to reduce the overall loss. This setting can be analogous to the additional output node used in multiple sequence alignment (MSA) predictions in AlphaFold2, the Nobel Prize-winning AI model for protein structure prediction, which leverages MSA knowledge to constrain the neural network’s parameter space and improve the accuracy of its main data-driven output: the 3D protein structure. With the help of such kind of a knowledge-embedded output node, during inference with a new instance, the model can leverage patterns learned from both the data and the knowledge-based rule to make predictions [37]. This approach in theory should be able to help reduce the risk of overfitting, particularly when the number of available data points is relatively limited. Finally, to enable relatively deeper learning, we trained both the KiNN and the corresponding FNN models for 3000 epochs.

Figure 6 shows the resulting R² distributions of different models. The mean of the R² distribution for the FNN trained on the subset (namely, FNN_subset) dropped to 0.64, compared to 0.78 delivered by the FNN trained on the entire data. It is also necessary to point out that a datapoint with a negative R² for the FNN_subset was removed from the analysis because it is considered an outlier (otherwise, the mean would have been even lower). The R² distribution of the FNN_subset model also exhibits a higher variance compared to the FNN, which may indicate increased uncertainty in the patterns learned from the reduced dataset. The KiNN results show that incorporating the additional domain-knowledge-based output head z increased the mean of the R² distribution to 0.72 while also reducing its variance notably. This clearly demonstrates that even the inclusion of a single piece of domain knowledge can enhance model performance under data-limited conditions. What is also interesting to observe is, even though “GDP per capita” is used as a predictor in the FNN_subset model, which potentially could have already captured the related domain knowledge, it did not hamper the added values of knowledge incorporation in the KiNN. This actually demonstrates that the factor encoding of knowledge into the loss function is more effective than encoding the knowledge into the layer of the neural network, which was also demonstrated when they numerically proved that encoding “structural chirality” knowledge into the attention layer of the AlphaFold2 model is less effective than encoding it into the loss function. Interestingly, although GDP per capita was already included as a predictor in the FNN_subset model, potentially capturing the relevant domain knowledge, it did not diminish the added value of explicitly incorporating that knowledge in the KiNN model. This suggests that encoding domain knowledge directly into the loss function is more effective than embedding it solely within the network architecture. This observation aligns with the findings of Jumper, Evans [37], who numerically demonstrated that encoding structural chirality as a constraint in the loss function of AlphaFold2 yielded much better results than embedding it into the model’s attention layer.

Hereby, it is important to emphasize again that the modelling analysis conducted in this study is intended only for demonstration purposes. We do not aim to provide a comprehensive analysis involving extensive hyperparameter optimization. Instead, we deliberately kept this case study simple to illustrate that incorporating domain knowledge can improve model performance in some situations. It is also important to note that the aim of this study is not to compare the proposed KiNN model with classical econometric or statistical models, such as panel data analysis. Instead, the focus is on demonstrating the added value of incorporating domain knowledge by comparing the performance of KiNNs with traditional FNNs. Future research could extend this work by evaluating KiNNs in comparison with classical models, particularly in contexts involving panel data, where capturing nonlinear patterns may provide additional insights.

4. Discussion

This project conducted research to use the national-level postharvest FL data registered by the FAO’s food balance sheets to predict the national (or even sub-national) level postharvest FL values with and without domain-knowledge incorporation.

For rice, the results show that the original neural network model achieves a high average R² value of about 95%, which shows its effectiveness in capturing data patterns by using the combined power of all predictors. The ablation study reveals the significance of certain predictors by evaluating performance drops when they are removed systematically each time. Predictors like “Year,” “Precipitation (mm),” “GDP per capita (USD),” and “Production (1000 t)” substantially impact model accuracy, which indicates that they can capture essential data patterns. The predictor “Year” reflects the time trend for postharvest FL changes due to the changing conditions (e.g., technology, awareness, etc.), which is not surprising. “GDP per capita” is related to the economic condition of a country or region that has a well-known relationship with postharvest FL (e.g., [20].) “Precipitation (mm)” is also expected to have a correlation with the rice loss, especially where the postharvest infrastructure is not good. The resulting high variances when eliminating the two predictors show the importance of them not only in delivering a high average prediction accuracy but also in reducing the variances of the performance and consolidate the patterns learned by the model. The relationship between the production of rice and its postharvest losses is weak, which may indicate that its major effect may already be captured by other predictors. In contrast, removing “FAO Region” and “Population (1000 persons)” shows minimal effect on model performance, likely because their information is already accounted for by other predictors or no relationship at all. For example, the FAO Region reflects the development stage of a country, which is also reflected by the “GDP per capita”. The geographical aspect of “FAO Region” can also be related to “Precipitation (mm)”. “The Population (1000 persons)” may simply not have any relationship with FL. Finally, “Protein Supply (g/capita/day)” negatively impacts performance, suggesting that it may add more noise compared to its added value in predictability. To verify the model’s generalizability, we also test the proposed method with wheat and apple. They deliver a lower prediction accuracy than the rice model but still quite good results, even though the uncertainty of the prediction also increases. It is also necessary to point out that, since the aim of this research is to demonstrate the potential of the proposed AI approach for proof of concept, we did not do much model fine tuning. In principle, for a deep learning model like the FNN, extensively fine tuning the model can usually improve the performance but requires substantially more computational time. In this sense, we have not explored the potential of the developed model to the full extent, which gives significant hope for future improvement.

In addition to the data-driven model, we also investigated the knowledge-based KiNN model to explore the added value of domain knowledge in guiding a more efficient learning process for the neural network. The results indicate that encoding domain knowledge into the loss function can partially mitigate the challenges posed by limited datasets. This highlights the significant potential of the KiNN to improve data utilization efficiency and reduce the risk of overfitting.

While the proposed approach demonstrates significant potential, several limitations must be acknowledged. Firstly, the FNN model is trained using the postharvest FL data derived from the FAO food balance sheets. These data are themselves outputs of a modelling framework that relies on primary data inputs and predefined assumptions. As a result, the FNN model is inherently bounded to the logic and mechanisms embedded in the FAO framework. This dependency restricts the model’s capacity to extrapolate beyond the limitations of the FAO methodology. For example, if the FAO framework oversimplifies or excludes certain regional or crop-specific factors, the AI model may inherit these limitations. This again highlights the importance of incorporating domain knowledge to enable the model to extrapolate beyond patterns that are learned purely from data. Secondly, the model’s performance is influenced by the selection of predictor variables, which, while capturing critical economic, demographic, climatic, and agricultural factors, may still omit some relevant factors to influence the postharvest food loss. For instance, variables such as infrastructure quality, supply chain efficiency, pest management practices, and storage technologies are omitted due to the poor data availability (too few datapoints to be included in the training dataset). The absence of these variables may reduce the model’s predictive accuracy in scenarios where such factors play a dominant role, even though they may have been indirectly considered by the currently used predictors. Thirdly, the model is trained on historical FAO food balance sheet data, which may limit its reliability in predicting long-term trends. Changes in agricultural practices, technological advancements, and policy interventions over time can introduce dynamics that the model may struggle to capture. This temporal aspect shows the challenge of applying the model to contexts involving rapidly evolving food systems or forecasting over a long period. Fourthly, this research is essentially a retrospective study rather than a prospective prediction, which may restrict its practical applications in policy making, particularly when addressing long-term, forward-looking issues. Finally, every robust machine learning model should be validated against primary data, which is lacking in this study due to its proof-of-concept nature.

5. Conclusions

Despite its limitations, we can still conclude that this study shows the potential of the proposed AI-based methodology for addressing data gaps in postharvest FL estimation. The FNN model and especially the KiNN model provide practical, cost-effective, and scalable alternatives to traditional data collection methods. This makes it potentially a valuable tool for enhancing food system management and policy development, particularly in scenarios where primary data are scarce or costly to obtain. This will especially benefit the countries that have a major problem with postharvest FL but do not have adequate resources to collect the primary data as the basis for potential interventions. Furthermore, incorporating domain knowledge into the data-driven framework through scientific machine learning is a promising approach to enhance the model’s performance, particularly when data are relatively limited. By integrating knowledge-based rules with neural networks, this approach can capture complex patterns, address data limitations, avoid overfitting, and improve the model’s applicability across diverse geographic and agricultural contexts. Finally, this research also contributes to promoting the circular economy by improving the prediction of postharvest losses, which thereby lays the foundation for potential waste valorization.