A FinTech-Aligned Optimization Framework for IoT-Enabled Smart Agriculture to Mitigate Greenhouse Gas Emissions

Abstract

With agriculture being the second biggest contributor to greenhouse gas (GHG) emissions through the excessive use of fertilizers, machinery, and inefficient farming practices, global efforts to reduce emissions have been intensified, opting for smarter, data-driven solutions. However, while machine learning (ML) offers powerful predictive capabilities, its black-box nature presents a challenge for trust and adoption, particularly when integrated with auditable financial technology (FinTech) principles. To address this gap, this work introduces a novel, explanation-focused GHG emission optimization framework for IoT-enabled smart agriculture that is both transparent and prescriptive, distinguishing itself from macro-level land-use solutions by focusing on optimizable management practices while aligning with core FinTech principles and pollutant stock market mechanisms. The framework employs a two-stage statistical methodology that first identifies distinct agricultural emission profiles from macro-level data, and then models these emissions by developing a cluster-oriented principal component regression (PCR) model, which outperforms simpler variants by approximately 35% on average across all clusters. This interpretable model then serves as the core of a FinTech-aligned optimization framework that combines cluster-oriented modeling knowledge with a sequential least squares quadratic programming (SLSQP) algorithm to minimize emission-related costs under a carbon pricing mechanism, showcasing forecasted cost reductions as high as 43.55%.

1. Introduction

Agriculture’s contribution to global greenhouse gas (GHG) emissions, ranking as the second largest source [1,2], presents a critical challenge to environmental sustainability and climate change mitigation efforts. In such applications, key GHG contributions can stem from various conventional farming practices and the management of livestock byproducts, including the excessive application of synthetic fertilizers, which leads to substantial nitrous oxide releases [3]; the energy consumption of and direct exhaust from farm machinery reliant on fossil fuels [4]; inefficient irrigation and land management techniques that can degrade soil carbon sinks [5]; and the anaerobic decomposition of organic matter, particularly manure from livestock, which releases considerable quantities of methane [6]. For this reason, global efforts towards more sustainable forms of agriculture have been intensified, opting for smart, data-driven solutions able to maintain productivity while significantly minimizing their GHG footprint, thereby addressing the pressing need for environmentally sound food production systems.

The existing literature has focused on mitigating emissions from a structural, land-use perspective, exploring macro-level strategies to enhance soil carbon sequestration through altered land management through afforestation initiatives [7] and by optimizing national land allocation for different agricultural activities through decision-support systems to spatially allocate land-use under policy and climate scenarios [8]. However, while these structural approaches are necessary for long-term sustainability, their implementation often depends on broad policy changes and faces socio-economic barriers, as they may not directly address granular operational inefficiencies or provide the immediate, market-based incentives required to promote adoption at the farm level [9,10].

In response to these challenges, the rapid evolution and convergence of advanced digital technologies, such as the Internet of Things (IoT), especially with the integration of advanced communication infrastructures (i.e., 6G-enabled IoT) with seamless service continuity [11] presents a transformative opportunity to address these agricultural sustainability challenges, enabling more sophisticated predictive analytics and machine learning (ML) applications [12]. Financial technology (FinTech) principles complement this technological capacity by introducing the economic dimension necessary for widespread adoption and the operationalization of sustainable goals [13]. Rather than relying solely on regulatory demands, FinTech offers tangible financial incentives for change, such as performance-based payments for verified emission reductions [14], access to preferential loan terms for investments in green farming technologies, and pollutant stock market mechanisms. Such strategies aim to actively reward emissions mitigation and promote proactive environmental management, thus being key in helping agricultural businesses align with increasingly important environmental, social, and governance (ESG) principles by making sustainable practices economically attractive, operationally efficient, and strategically advantageous for the modern farmer.

However, while advanced ML approaches have demonstrated remarkable predictive accuracy, their inherent black-box nature often poses a significant challenge to practical implementation and stakeholder trust. In a domain where recommendations directly impact operational practices and financial outcomes, the inability to understand why a model suggests a specific action can lead to skepticism and low adoption rates. This opacity is particularly problematic when integrating with FinTech mechanisms; for an IoT system of financial incentives and penalties to be considered fair and effective, its decision-making logic must be transparent, auditable, and easily communicated [15].

To fill this gap, this study proposes a novel FinTech-aligned GHG emission optimization framework specifically designed to integrate seamlessly with IoT-enabled agricultural systems, supporting both economic viability and environmental sustainability and facing the current challenges in smart agriculture. By incorporating key environmental factors, such as soil health, livestock management, and energy emission metrics, alongside socio-economic parameters, this proposed framework uses statistical modeling to dynamically adjust farming recommendations, expanding upon previous research on integrating 6G-IoT in smart agriculture for GHG emission mitigation [2]. Inspired by pollutant stock market mechanisms, ways to effectively integrate market-based strategies within agriculture were also explored in order to encourage emissions reductions and align with broader ESG principles. Unlike macro-level land-use studies, this framework provides transparent, market-aligned recommendations directly to stakeholders, and in contrast to black-box ML models, it ensures its decision-making logic is both auditable and interpretable.

The contribution of this work is twofold and mainly connected to the two core interlocked methodologies, namely GHG emission modeling and the FinTech-aligned optimization framework, as seen in the process flowchart depicted in Figure 1, focusing specifically on the Edge/Cloud layer of the IoT infrastructure rather than the physical data acquisition side. First, this work begins with the exploration and modeling of key GHG emission drivers by combining principal component analysis (PCA) with k-Means clustering to reduce the dimensionality of the original feature space and define emission profiles, then implementing a cluster-oriented principal component regression (PCR) to capture any underlying linear data relationships. This extracted knowledge is then combined with pollutant stock market mechanisms and passed on to develop the proposed FinTech-aligned optimization framework centered around interpretability, providing a transparent and economically grounded approach to agricultural emissions mitigation unlike other existing black-box ML approaches.

The remainder of this paper is structured as follows. In Section 2, the technological background for this study is offered, providing an overview of the key GHG emission contributors in smart farming and describing similar frameworks from the recent literature. The contribution of this work is also highlighted at the end of this section. Section 3 and Section 4 are related to the proposed emission modeling and optimization framework methodologies, respectively, while Section 5 provides some initial experimental results in each corresponding subsection. Finally, Section 6 provides a comprehensive discussion on the proposed decarbonization strategies and the way they can implemented into abatement costs for smart agricultural applications, and Section 7 concludes the paper by summarizing important insights and providing future steps.

2. Background and Research Contribution

To contextualize the proposed framework, this section first presents a basic background on agricultural GHG emissions, identifying the primary contributors and their sources. With this context established, a comprehensive review of the most recent literature is then provided, covering data-driven optimization and modeling frameworks for agricultural emission mitigation to identify the existing gaps and highlight the contribution of this work.

2.1. GHG Emissions and Primary Contributors in Agriculture

Agricultural systems are widely recognized as key contributors to global GHG emissions, accounting for approximately 13–21% of total anthropogenic emissions, depending on whether land-use change is considered [16]. These emissions are not homogenous but comprise three principal gases—nitrous oxide (N₂O), methane (CH₄), and carbon dioxide (CO₂)—each characterized by distinct sources, mechanisms of formation, and climatic impacts.

Methane, a highly potent GHG, is primarily emitted through biological processes within agricultural contexts, with enteric fermentation in ruminant livestock, such as cattle, sheep, and goats, producing considerable amounts of methane as a metabolic byproduct [17]. Rice cultivation represents another major contributor, with the sustained flooding of paddy fields fostering anaerobic soil conditions, which in turn favor methanogenic microbial activity and consequently elevate methane emissions [18,19]. Additionally, the anaerobic decomposition of livestock manure during storage further amplifies methane emissions, highlighting the importance of improved manure management strategies as a key mitigation practice.

Nitrous oxide emissions, on the other hand, are strongly associated with nitrogen cycling in agricultural soils, a practice that initiates a series of microbial transformations, including nitrification (i.e., oxidation of ammonia to nitrate under aerobic conditions) and denitrification (i.e., reduction of nitrate to nitrogenous gases, including N₂O, under anaerobic conditions) [20,21]. The magnitude of nitrous oxide emissions is influenced by multiple biophysical and management factors, including fertilizer type and application rate, soil texture and composition, moisture regimes, and ambient temperature. Similar to methane, manure management contributes to nitrous oxide emissions, particularly through the storage and application of organic nitrogen sources on croplands.

Finally, carbon dioxide emissions in agriculture largely result from fossil fuel combustion and land-use change, with the direct utilization of diesel and gasoline to operate farm machinery, irrigation systems, and other on-site equipment serving as a primary source of carbon dioxide [22]. Indirect emissions also result from the energy required for the production and distribution of agricultural inputs and outputs, as well as from heating and cooling systems used in controlled-environment agriculture. Furthermore, the conversion of forests and other carbon-rich ecosystems to agricultural land leads to substantial carbon dioxide emissions due to biomass burning and soil carbon losses [23]. However, while carbon dioxide has the longest atmospheric lifetime among the three gases, methane and nitrous oxide exhibit far higher global warming potentials (GWPs), estimated at 28–36 and 265–298 times greater than carbon dioxide, respectively, over a 100-year time horizon [24]. This mismatch highlights the necessity of addressing all three GHGs for effective climate mitigation strategies within the agricultural sector.

2.2. Related Work

Research on GHG emission mitigation in agriculture has grown considerably in recent years, particularly through the development of data-driven and optimization-based decision frameworks that aim to balance productivity and sustainability goals by minimizing emissions associated with critical agricultural activities, such as crop cultivation, livestock management, energy use, and input application. Optimization frameworks are increasingly employed to enhance both environmental and economic outcomes through precise modeling, predictive control, and intelligent decision support.

In this context, several studies have applied linear and multi-objective programming models to optimize resource allocation and farming structure. These models typically seek to balance economic returns with environmental impacts, often using multi-objective linear programming to determine the optimal scale of different industries in a circular agriculture system that benefits from internal resource flows [25]. Similarly, sophisticated setpoint optimization strategies have been developed for greenhouse environments, where environmental control variables (temperature and CO₂ concentration) are adjusted to maintain optimal growing conditions. These strategies often incorporate global optimization techniques, such as surrogate-based methods [26] or optimal control schemes that account for spatial distributions [27], to derive climate control policies that jointly consider crop yield and energy efficiency.

Reinforcement learning (RL) has also been widely adopted in smart farming systems, particularly for adaptive climate control and optimizing the use of agricultural inputs, having the ability to autonomously learn optimal strategies for fertilizer and irrigation application by interacting with biophysical simulators [28]. In greenhouse applications, RL is used to create adaptive controllers that learn from real-time data, often collected via IoT devices, to minimize energy consumption while maintaining desired climate setpoints [29]. In certain applications, RL frameworks are integrated into model predictive control architectures that continuously adjust system dynamics and constraints online, allowing for improved performance under uncertain or changing conditions [30].

Other approaches emphasize the integration of emission modeling into comprehensive decision support systems and analytical platforms. Some studies have built platforms based on Intergovernmental Panel on Climate Change (IPCC) methodologies that use IoT to collect comprehensive farm data from soil health to livestock activity, providing farmers with clear visualizations to manage GHG emissions in mixed farming systems [31]. Other research focuses on identifying the primary drivers of emissions through large-scale data analysis. In this context, studies have used comprehensive Food and Agriculture Organization (FAO) datasets and statistical methods, including PCA and recursive feature elimination (RFE), to identify the key factors influencing emission patterns [32]. To enhance predictive accuracy and interpretability, hybrid ML models combining long short-term memory (LSTM) with random forests have also been used to forecast GHG emissions and quantify the contribution of each input factor, informing targeted mitigation strategies [33].

2.3. Problem Definition and Research Contribution

While prior studies have significantly contributed to agricultural GHG management, they primarily fall into two distinct methodological directions. The first involves prediction- and control-oriented approaches, such as RL-based or model predictive control frameworks [28,30], which, although highly effective at dynamic optimization, often function as “black boxes”, sacrificing model transparency and failing to provide deeper insights into the factors that drive their recommendations. On the other hand, the second direction consists of explanation-focused analytical models using statistical- and ML-based techniques to identify and rank key emission drivers [32,33]. These approaches, while offering valuable diagnostic insights compared to the previous direction, are typically descriptive in nature, lacking an integrated mechanism to translate those insights into an optimized, prescriptive course of action for IoT-enabled systems, highlighting the need for a framework that can combine prescriptive optimization with inherent model interpretability.

For this reason, the proposed FinTech-aligned optimization framework is specifically designed to fill this gap by prioritizing an explanation-focused and transparent methodology with the following characteristics:

• Integrated diagnostics and predictions: The proposed framework implements a two-stage approach that first models the underlying GHG emission drivers from IoT sensor networks and then uses this embedded knowledge to perform the optimization.
• Interpretable recommendations: The framework provides direct, quantifiable coefficients that link specific management variables to emission outcomes.
• Cluster-oriented and profile-specific optimizations: The framework defines distinct emission profiles before optimization, offering a more tailored approach specific to the unique context of different IoT-enabled agricultural systems.

3. Greenhouse Gas Emission Modeling

In this section, the complete GHG emission modeling methodology is analyzed step by step, starting from data collection and preprocessing and leading to the development and implementation of the proposed cluster-oriented regression model.

3.1. Dataset Description and Preprocessing

The original dataset utilized in this work was obtained from the Food and Agriculture Organization Corporate Statistical (FAOSTAT) database [34], which is known for its extensive and long-term global coverage. For the purposes of this study, data from all European countries from 1961 to 2022 were selected, focusing on three key domains: GHG emissions from crops, livestock, and energy use. It should be noted that, given the dataset’s nature as a high-level statistical compilation, it does not include granular information from individual farm operations. Instead, all emissions are primarily Tier 1 estimates calculated using IPCC methodologies that rely on combining activity data (e.g., livestock populations) with default emission factors [35]. For this reason, this work focuses on modeling the relationships between these macro-level activity drivers and their corresponding emission outcomes.

To create the final emission dataset, all three of the aforementioned separate source files were combined into one unified dataset, which was further streamlined by removing all descriptive metadata columns, as they provided no further value for modeling. In addition, all reported values were standardized into a consistent unit of megatonnes (Mt) to ensure numerical comparability across different emission types, while activity data such as agricultural area (ha) and energy use (TJ) were maintained in their original units for subsequent analysis.

Following the initial cleaning, the main data preprocessing methodology involved aggregating the highly granular data into a set of interpretable, high-level features suitable for modeling by developing a mapping logic that combined related sub-categories into broader ones. This aggregated dataset was then pivoted from a long to a wide format, creating a final matrix where each row represents a unique country–year observation. Finally, total CO₂-equivalent emissions were calculated as the target variable for the model by applying the 100-year GWP factors of 298 for N₂O and 28 for CH₄ to the relevant emission columns [36]. Any remaining missing values, which typically indicated the absence of a reported activity for a given year, were filled with zero [35].

3.2. Principal Component Analysis and Interpretation

Following the data preprocessing methodology, PCA was implemented on the final scaled FAOSTAT dataset to reduce dimensionality and identify the most influential latent variables impacting agricultural GHG emissions by compressing the information into a smaller set of uncorrelated components that capture the majority of the data’s variance while limiting redundancies [37]. To establish the optimal number of principal components (PCs) to retain, a scree plot was produced (Figure 2), representing the explained variance of each component in descending order (Figure 2a) [38]. Based on this plot, and following the elbow method [39], five PCs were selected as the optimal solution, offering a cumulative explained variance of approximately 98% (Figure 2b), ensuring that the reduced feature set effectively represented the original data. It should be noted that the minimal variance explained by subsequent components (i.e., top-6th PC) suggests that they primarily capture noise rather than meaningful patterns [40]; thus, we excluded the 6th principal component from our final selection.

Figure 3 depicts the significance of each original feature’s influence on the five selected PCs, with the PC loadings showcasing the variables that contribute most to explaining the variance in GHG emissions. For instance, PC1 is mainly influenced by a combination of features with high positive loadings, including the number of animals (i.e., Livestock_Stock), as well as nitrous oxide emissions from both the decomposition of leftover straw returned to the soil and energy usage (i.e., Crops_Residues_N2O and Energy_N2O), suggesting that PC1, which explains the largest portion of the variance, primarily captures the overall scale and intensity of agricultural operations. On the other hand, PC2 reveals a different set of dominant variables related to specific types of agricultural systems. It is characterized by high positive loadings from methane emissions generated in the anaerobic conditions of flooded rice paddies (i.e., Crops_RiceCult_CH4) and the total land area under cultivation (i.e., Crops_Area). This is contrasted by strong negative loadings for livestock-related features, mainly methane released from the digestive process of ruminants (i.e., Livestock_Enteric_CH4), indicating that this component distinguishes between cropping-dominant and livestock-dominant systems. Similarly, PC3 highlights a contrast between operations driven by on-farm energy consumption versus those driven by biological livestock processes, making it strongly influenced by carbon dioxide emissions from the combustion of fossil fuels (i.e., Energy_CO2), with negative loadings for biological emissions like methane from stored manure (i.e., Manure_Mgmt_CH4) and animal digestion (i.e., Livestock_Enteric_CH4). Other later components, such as PC4 and PC5, capture more specialized variations, with PC4 being strongly defined by burning crop residues (i.e., Crops_BurningResidues_CH4) and PC5 by methane emissions from on-farm energy use (i.e., Energy_CH4), isolating the effects of specific management practices.

Finally, the PC value for each country–year observation was calculated as the weighted sum of the standardized variable values, with weights corresponding to their respective loadings [37], as analyzed in Equation (1) as follows:

$${\mathrm{PC}}_j=\sum _{i=1}^n({\mathrm{loading}}_{ij}\times {\mathrm{variable}}_i)$$

(1)

where ${\mathrm{PC}}_j$ represents the score of a single data point (e.g., one country–year observation) on the j-th principal component, n is the total number of original features, ${\mathrm{loading}}_{ij}$ is the weight of the i-th original variable on the j-th principal component, and ${\mathrm{variable}}_i$ is the standardized value of the i-th original variable for that single data point.

The heatmap in Figure 4 illustrates the significance of each original feature’s influence on the five most significant PCs, with dark and light blue colors representing strong positive and negative influence, respectively, and in-between hues indicating smaller influence on each principal component. For instance, PC1 exhibits uniformly positive loadings across nearly all features, with the highest values exhibited for the number of animals, nitrous oxide emissions from crops, and a cluster of energy, manure, and enteric fermentation features, while PC2 clearly distinguishes between different farming systems, showcasing very high positive loading values on crop emissions and a very strong negative correlation with livestock-related features. PC3 highlights a trade-off between technological and biological emission sources, showcasing strong positive loadings for energy-related features and strong negative loadings for manure and enteric fermentation ones, differentiating systems based on whether their emission profile is primarily driven by fossil fuel consumption or by biological processes inherent to livestock management. Finally, PC4 and PC5 capture more specific variations associated with managing crop residues and distinguishing systems based on the type of energy used, respectively.

3.3. Clustering Analysis and Emission Profiles Definition

Following the PCA, a clustering analysis was also performed to classify the country–year observations into distinct profiles based on their agricultural activity and emission patterns. By classifying the original national-level observations into distinct, more homogeneous profiles based on their underlying agricultural activity and emission patterns, this clustering step is intended to partially mitigate the inherent limitations of macro-level data in capturing field management heterogeneity [41].

To ensure the clustering was informed by both the emitters and the outcomes, the analysis was conducted on a dataset combining the five principal components with the standardized total GHG emissions. To determine the optimal number of clusters, a range of k values was evaluated for which the data were reduced to the PCA-reduced dimensional space of the optimal principal components, and k-Means clustering was applied to assign data points to each cluster [42,43,44].

As showcased in both the 3D and 4D plots in Figure 5, the optimal number of clusters was found to be k = 4, determined using the silhouette score [45,46], creating four clusters with low overlap. The scatter plot in Figure 5a visualizes the four clusters based on the first two principal components and their total GHG emissions (i.e., CO₂ equivalents), showcasing that clusters 0 and 1 are grouped together at the bottom of the plot, confirming their lower total emissions. In contrast, clusters 2 and 3 are depicted higher, indicating respective larger carbon footprints. Figure 5b further illustrates this separation within the space of the first three principal components, showing that each of the four clusters forms a dense and distinct group with minimal overlap.

To further interpret these four clusters, a reverse PCA was also implemented to reconstruct the original FAOSTAT emission dataset using only the most significant principal components. In Figure 6 are depicted the cluster centroids projected back to the original emission feature space, displaying the standardized mean value for each feature, allowing for a direct comparison of the characteristics that define each cluster, offering highly interpretable results compared to other clustering methods [47].

As seen from the bar plot, the four clusters are mainly differentiated by their dominant agricultural practices, input intensity, and resulting emission profiles, with cluster 0 (“CH4-Heavy Cropping”) representing systems with exceptionally high methane energy emissions (∼1.7 stds above the mean) but low livestock activity, cluster 1 (“Low-Input Traditional”) representing low-intensity operations with values for most emission drivers significantly below the mean (i.e., ∼1.4 stds below the mean for energy and manure inputs), cluster 2 (“High-Emission Livestock”) signifying specialized livestock systems with extremely high values for number of animals and related emissions (∼1.5 stds above the mean), and cluster 3 (“Intensive Cropping and Manure”) representing highly intensive, mixed systems with the highest levels of nitrous oxide emissions from both crops and manure (both ∼1.6 stds above the mean).

Based on the above analysis, it is evident that cluster 2 showcases the highest standardized mean value for total GHG emissions (∼1.4 stds above the mean), confirming that extremely large, specialized livestock operations lead to the largest overall carbon footprint. On the other hand, cluster 1 exhibits the lowest standardized mean value for total emissions (∼1.3 stds below the mean), highlighting that traditional, low-input systems have a significantly smaller environmental impact and providing a clear benchmark for low-emission agriculture. Finally, cluster 3 represents a highly intensive, mixed-system profile that is the second-largest overall emitter, while cluster 0 demonstrates that a specialized cropping system can still maintain below-average total emissions, highlighting that while livestock specialization drives the highest carbon footprint, other forms of specialization do not necessarily lead to the same outcome.

To further validate the agricultural logic of the aforementioned emission profiles, their geographical distribution was analyzed by assigning each country to its most frequently occurring cluster during the recent dataset decade (i.e., years 2012–2022). As seen in Figure 7, the resulting spatial patterns strongly align with known agricultural structures. For instance, cluster 0 is uniquely represented by Germany, a finding consistent with its highly mechanized and energy-intensive cropping sector [48], while cluster 2 groups major agricultural producers, including the United Kingdom, France, and Russia, aligning with their large-scale, intensive livestock operations, known to be significant drivers of GHG emissions [49,50]. Cluster 3 is concentrated in Mediterranean countries (i.e., Spain and Italy), reflecting their high-input agricultural systems that combine intensive horticulture with significant manure production [51]. Finally, cluster 1 covers all remaining countries, mainly in Northern and Eastern Europe, generally corresponding to regions with less intensive or more traditional agricultural practices compared to the other highly specialized profiles [52], further validating the provided clustering results.

3.4. Greenhouse Gas Emission Modeling

To closely fit Europe’s GHG emission data, also using the collected information from both the PCA and the clustering analysis, a cluster-oriented PCR model was implemented. Unlike datasets derived from complex physical processes, Tier 1 emission estimates are based on a linear accounting framework where emissions are calculated by multiplying activity data (i.e., number of cattle or amount of fertilizer applied) by a standardized emission factor. Given that the principal components are, by definition, linear combinations of these underlying activity drivers, the relationship between the PC scores and the total GHG emissions is also inherently linear, making PCR a suitable approach for modeling these specific FAOSTAT emissions.

The proposed emission modeling methodology extends the core multiple linear regression (MLR) principles [53], training a separate MLR model for each of the four identified clusters and using the established top five principal components as estimators, as explained in the following Equation (2).

$$g{\left(E\left(Y\right)\right)}^{\left(k\right)}={\beta }_0^{\left(k\right)}+\sum _{j=1}^p{\beta }_j^{\left(k\right)}{\mathrm{PC}}_j$$

(2)

Here, $g\left(\right)$ is the link function that relates the expected value of the response variable Y to the estimators (in CO₂-equivalents) belonging to cluster k, ${\beta }_0$ is the intercept for cluster k, and ${\mathrm{PC}}_j$ denotes the j-th principal component from the PCA. The regression coefficient ${\beta }_j^{\left(k\right)}$ is applied to each of the top five principal components, quantifying the magnitude and direction of the relationship between that specific latent emitter and the total GHG emissions for that cluster’s profile.

3.5. Evaluation Metrics for Modeling

To assess the fitting accuracy of the proposed cluster-oriented PCR model, two widely-accepted regression metrics were employed. The primary metric, cross-validated root mean squared error (RMSE), was used to quantify the average magnitude of prediction error in the original units of total GHG emissions. However, the scale-dependent nature of RMSE makes direct performance comparisons between clusters challenging, as each agricultural profile exhibits a distinct range of emission values. To overcome this limitation and facilitate a standardized evaluation, the normalized cross-validated RMSE (nRMSE) was also calculated, scaling the error by the range of the actual target values ($max\left(y\right)-min\left(y\right)$) [54], thereby expressing the error as a proportion of the data’s spread and allowing for a fair, scale-independent comparison of model effectiveness across the diverse clusters.

4. Greenhouse Gas Emission Optimization

Following the emission data modeling, this section describes the proposed FinTech-aligned GHG emission optimization methodology, starting from the optimization problem and constraint definition and leading to the development of the actual framework.

4.1. Optimization Problem and Constraint Definition

Following the establishment of the GHG emission driver modeling framework, this descriptive analysis pivots to a prescriptive role, creating an optimization framework capable of identifying actionable and economically viable mitigation recommendations by translating continuous data streams into optimized control actions. However, the core challenge is not just to minimize emissions, but to do so in a transparent and trustworthy way, aligned with the economic realities of agricultural producers and policymakers.

To achieve this, our proposed approach operationalized the environmental target by utilizing market-based mechanisms inspired by FinTech principles, such as carbon taxes [55]. By introducing an external carbon price, the optimization problem is transformed from an abstract environmental goal into a tangible financial one, minimizing the total economic cost associated with a country’s forecasted GHG emissions. For the purposes of this study, a baseline price of EUR 85 per tonne of CO₂ equivalents was selected, reflecting the approximate average market price within the EU Emissions Trading System (EU ETS) during 2023 [56]. This financial objective aligns with established economic modeling practices, which frequently use carbon pricing to construct marginal abatement cost curves and quantify the trade-offs between economic cost and emission reduction levels in the agricultural sector [57].

However, this financial framing also requires model transparency. For an IoT system of financial incentives or penalties to be considered fair and effective, its decision-making logic must be fully auditable and comprehensible to stakeholders. This presents a significant challenge for black-box optimization models, as their lack of interpretability might limit trust and practical adoption in high-stakes domains [58]. Our framework is explicitly designed to overcome this limitation. Instead of optimizing on hundreds of granular and often correlated agricultural inputs, the decision variables are the scores of the pre-selected optimal principal components, which, thanks to their high-level and interpretable strategic directions (e.g., “overall agricultural intensity” or “crop vs. livestock specialization”), make the optimization output an inherently understandable recommendation.

Finally, the optimization is subject to realistic constraints that reflect practical abatement potential. It is recognized that the capacity for agricultural systems to change is limited by socio-economic, technological, and biophysical factors in the short term [59]. To model this, our proposed framework introduced limits on the optimized PC scores, allowing them to deviate by a maximum of 20% from their forecasted actual trajectory, ensuring that the generated recommendations are not only mathematically optimal but also grounded in a feasible scope.

4.2. FinTech-Aligned GHG Emission Optimization

The proposed optimization framework was implemented as a sequential pipeline using Python’s (version 3.11.9) Scikit-learn library for predictive modeling and SciPy for numerical optimization. Since the actual data from the FAOSTAT database included only years from 1961 to 2022, the first step was to forecast the actual emission data for 2025, which served as the baseline for optimization. For each country, a simple linear regression model [60] was fitted to the historical time-series of its five PC scores to extrapolate the trend to 2025, given the inherently linear nature of both the initial FAOSTAT data and the calculated PC scores [61].

Once the baseline PC scores were forecasted, each country’s future agricultural profile was classified using the pre-trained k-Means model from the clustering step, assigning each country to one of the four emission profiles, thereby determining which of the cluster-specific PCR models would be used to model its GHG emissions during optimization.

Finally, the optimization function was executed for each country with the goal of minimizing the total financial cost of emissions, calculated as the predicted GHG output multiplied by an external carbon price. The five forecasted PC scores are the decision variables for this minimization, and the optimal solution is obtained using a sequential least squares quadratic programming (SLSQP) algorithm [62], as seen in Equation (3).

$$\underset{\mathrm{PC}\prime }{min}c\left(\mathrm{PC}\prime \right)=P_{\mathrm{carbon}}\times ({\beta }_0^{\left(k\right)}+\sum _{j=1}^p{\beta }_j^{\left(k\right)}{\mathrm{PC}\prime }_j)$$

(3)

where ${\mathrm{PC}\prime }_j$ relates to the optimized PC scores, $c\left(\mathrm{PC}\prime \right)$ is the objective function representing the total financial cost of emissions, $P_{\mathrm{carbon}}$ is the external carbon price in EUR per tonne, and ${\beta }_0^{\left(k\right)}+{\sum }_{j=1}^p{\beta }_j^{\left(k\right)}{\mathrm{PC}\prime }_j$ is the aforementioned cluster-oriented PCR model using the optimized PC scores.

To ensure the recommendations are achievable, the optimization is bounded, regulating the optimized PC scores to a feasible range of ±20% of their forecasted baseline values, acknowledging that agricultural systems are limited in their capacity for rapid structural change [59,63]. This specific ±20% threshold was selected to represent a plausible level of short-term mitigation, a choice directly informed by the relevant literature, including comprehensive studies on agricultural non-CO₂ mitigation, suggesting that cost-effective structural changes in livestock and crop systems could deliver global emission reductions of up to 21% [57]. This proposed method of using a bounded step is also aligned with agricultural economic studies on marginal abatement cost curves, which often model mitigation potential in discrete, achievable steps rather than as an unbounded continuum [64]. The final output of the optimization model is a new optimized set of PC scores that represents the recommended mitigation strategy for each country.

4.3. Evaluation Metrics for Optimization

To evaluate the effectiveness of the proposed optimization framework, two primary performance metrics were defined to reflect both the environmental and economic dimensions of the problem. Emission reduction is the most direct one, quantifying the absolute environmental benefit of the optimized strategy, calculated as the difference between the forecasted actual emissions and the final optimized emissions and expressed in Mt of CO₂-eq. [59]. While this metric provides a clear measure of the potential climate impact, it does not in itself account for the economic feasibility of achieving such reductions. For this reason, cost reduction, defined as the percentage decrease in the total financial cost of emissions between the baseline and optimized scenarios, was also calculated to assess the financial viability and efficiency of the recommendations. By expressing the benefit in relative economic terms, this metric aligns with economic analyses that use carbon pricing to evaluate the cost-effectiveness of abatement strategies [57] and provides a scale-independent measure of the framework’s ability to generate financially attractive mitigation recommendations.

5. Experimental Results

In this section, both the fitting and the optimization results from the complete FinTech-aligned optimization framework are offered, analyzed in each corresponding subsection.

5.1. Data Fitting Results

To evaluate the cluster-oriented PCR model’s performance, five-fold cross-validation was performed to provide a stable and reliable estimate of model generalization performance [47].

Table 1. Fitting results of the cluster-oriented principal component regression modeling approach for different numbers of principal components within each cluster.

	Cluster-Oriented Principal Component Regression
	Cluster 0		Cluster 1		Cluster 2		Cluster 3
PCs	RMSE	nRMSE	RMSE	nRMSE	RMSE	nRMSE	RMSE	nRMSE
Top 1	6.043	0.068	4.072	0.078	13.952	0.058	3.224	0.202
Top 2	6.332	0.071	4.053	0.077	9.773	0.040	3.140	0.197
Top 3	1.112	0.012	2.829	0.054	5.859	0.024	1.209	0.076
Top 4	1.088	0.012	0.991	0.019	2.459	0.010	1.140	0.071
Top 5	0.622	0.007	0.936	0.018	1.965	0.008	1.080	0.068

showcases the model’s fitting accuracy as the number of PCs used as features was incrementally increased from one to five. As seen from the table, there is a consistent improvement in model performance with the inclusion of additional components across all clusters. While the first two PCs provide a coarse approximation, the most substantial gains in accuracy occur with the inclusion of the third, fourth, and fifth components. For instance, in cluster 0, the nRMSE value dropped dramatically from 7.1% with two PCs to just 0.7% with five, while in cluster 3, the nRMSE value improved from 19.7% to 6.8%. Since the lowest error values for every cluster were achieved when using all five PCs, this configuration was selected as optimal, confirming that even the later components, while explaining less overall variance, contain valuable information to accurately model GHG emissions within each specific profile.

To further validate our proposed modeling approach, a comparative analysis against its simpler generalized variant was conducted. As seen in

Table 2. Fitting results of the proposed cluster-oriented principal component regression approach compared to the simpler generalized variant.

	Generalized PCR		Cluster-Oriented PCR		Accuracy Improvement
Cluster	RMSE	nRMSE	RMSE	nRMSE
C0—“CH4-Heavy Cropping”	1.282	0.014	0.622	0.007	0.515
C1—“Low-Input Traditional”	1.104	0.021	0.936	0.018	0.152
C2—“High-Emission Livestock”	2.354	0.010	1.965	0.008	0.165
C3—“Intensive Cropping and Manure”	2.438	0.153	1.080	0.068	0.557

, the performance of the proposed cluster-oriented PCR was benchmarked against a simpler, generalized PCR variant trained on the entire PCA-reduced dataset, and the results demonstrated the clear superiority of the cluster-specific approach across all four agricultural profiles. The most significant accuracy improvements were observed for clusters 0 and 3, where the RMSE values dropped by 51.5% and 55.7%, respectively. Clusters 1 and 2 also showed notable performance gains of 15.2% and 16.5%, confirming that segmenting the data into distinct emission profiles and developing a tailored modeling approach allows us to capture the unique underlying relationships more effectively, leading to a more accurate and reliable modeling foundation for the subsequent optimization.

Based on the result evaluation, notable differences in modeling accuracy across clusters were revealed. More specifically, clusters 0 and 2 consistently achieved the lowest nRMSE scores, indicating that their emission patterns, though different in scale, are highly predictable once their specialized nature is considered. In contrast, cluster 3 showcases the highest error overall, suggesting that the complex interactions between multiple high-emission activities in these mixed systems introduce greater variability, making their outcomes inherently harder to model with a linear approach. Finally, cluster 1 lies between these extremes, exhibiting moderate and stable accuracy.

These results are further supported by the proposed PCR model’s residual analysis, depicted in Figure 8, conducted for each of the four clusters separately. As seen in the plots, the residuals for clusters 0 and 2 are distributed fairly randomly, with no distinct patterns, and their tightly centered estimations around the $y=x$ line, also supported by their very low nRMSE values of 0.7% and 0.8%, respectively, verify the model’s high fitting power [65]. For cluster 3, which represents more complex mixed systems, the “Estimated vs. Actual” plot confirms an unbiased linear relationship, although with slightly higher variance than the previous clusters, while the random scatter in the residual plot supports the model’s acceptable and accurate representation. Finally, for cluster 1, while the model also demonstrates high overall fitting accuracy, both plots reveal a subtle second-order non-linear effect, visible through the slight S-curve in the “Estimated vs. Actual” plot and the inverted-U shape in the residuals [60]. However, while a non-linear model could potentially capture this minor effect, the extremely low nRMSE value of 0.018 confirms that the linear model already provides a very strong and practically useful approximation for this cluster.

5.2. Emission Optimization Results

Following the evaluation of the cluster-oriented modeling method, the emission optimization results from the final framework are offered for the top European GHG emitters. As seen from

Table 3. Optimization results of the proposed FinTech-aligned framework for each of the top European GHG emitters.

Country	Cluster ID	Emissions (Mt)	Optimized Emissions (Mt)	Emission Reduction (Mt)	Cost Reduction (%)
France	2	97	97	0	0.00
Russia	2	74	42	32	43.55
Germany	0	62	62	0	0.00
United Kingdom	1	62	53	9	13.43
Spain	3	54	41	13	23.42
Italy	3	44	26	18	41.33
Poland	1	33	33	0	0.00
Netherlands	1	31	21	10	31.56
Ireland	1	27	24	3	10.58
Belarus	1	20	20	0	0.00
Romania	1	19	16	3	15.94
Belgium	1	12	10	2	18.69

, the most substantial opportunities for cost-effective emission reduction were identified in Russia and Italy. Russia, classified under cluster 2, achieved a significant 43.55% reduction in emission-related costs, corresponding to an absolute decrease of 32 Mt CO₂-eq. Similarly, Italy, belonging to cluster 3, demonstrated a 41.33% cost reduction, equivalent to 18 Mt CO₂-eq., highlighting the framework’s effectiveness at identifying inefficiencies within large-scale, high-emission systems where significant improvements are possible within the defined constraints.

The zero-emission-reduction results for France, Germany, Poland, and Belarus do not indicate algorithm failure, but rather suggest that their forecasted actual profiles for 2025 were already at or near an optimal state within the imposed ±20% constraint boundary. Since this is a convex optimization problem, any locally optimal point is also globally optimal [66]; thus, a zero-reduction outcome signifies that for the forecasted 2025 profile, any permissible short-term adjustment would lead to an increase, not a decrease, in emission-related costs. As a result, for these countries, any further significant mitigation would likely require more substantial, structural changes beyond the scope of the short-term abatement potential modeled in this framework.

The remaining countries, primarily classified under cluster 1, showed moderate to significant potential for improvement. The Netherlands, for instance, achieved a substantial 31.56% cost reduction, while the United Kingdom, Romania, and Belgium demonstrated reductions ranging from 13% to 19%, indicating that even within lower-intensity systems, the framework can effectively identify and recommend valuable optimization recommendations. The re-classification of the United Kingdom into cluster 1 for its 2025 forecast, despite its historical data often aligning with the “High-Emission Livestock” profile of cluster 2, indicates that the time-series extrapolation of its PC scores captures a significant long-term structural trend away from specialized, high-intensity livestock operations and towards a more diversified emission profile relative to its European peers. Overall, the outcomes highlight that the potential for emission mitigation is highly context-dependent, and the proposed profile-specific optimization successfully identifies where the most significant and financially attractive abatement opportunities exist. Figure 9 showcases the emission reduction potential of all European countries included in the initial FAOSTAT dataset, as extracted from the optimization function.

6. Emission Control via Market-Informed Optimization: A European Case Study

Sensitivity Analysis on Carbon Price

To evaluate the economic implications of the proposed optimization framework under different policy scenarios, a sensitivity analysis [67] was conducted on the external carbon price to assess the way that the financial savings from the identified emission reductions scale with varying carbon pricing levels. For the purposes of this study, three scenarios were modeled, assuming a “low price” of EUR 50 per tonne of CO₂-eq. [56], a “baseline price” of EUR 85 per tonne, which is used in the primary analysis, and a “high price” of EUR 120 per tonne [68]. The results for the ten European countries with the highest absolute emission reduction potential are presented in

Table 4. Sensitivity analysis results of forecasted financial savings for the top 10 European countries with the highest GHG emission reduction potential for 2025 under varying carbon pricing scenarios.

Country	Emission Reduction (Mt)	Low-Price Savings (Eur)	Baseline Savings (Eur)	High-Price Savings (Eur)
Russia	32	1.6 M	2.7 M	3.9 M
Italy	18	0.9 M	1.6 M	2.2 M
Spain	13	0.6 M	1.1 M	1.5 M
Netherlands	10	0.5 M	0.8 M	1.2 M
United Kingdom	8	0.4 M	0.7 M	1.0 M
Andorra	4	0.2 M	0.3 M	0.5 M
Cyprus	4	0.2 M	0.3 M	0.5 M
Luxembourg	4	0.2 M	0.3 M	0.5 M
Armenia	4	0.2 M	0.3 M	0.4 M
Slovenia	4	0.2 M	0.3 M	0.4 M

, which reveals a direct correlation between the carbon price and the magnitude of financial savings.

As seen in the table, countries with the largest absolute emission reduction potential, such as Russia (i.e., 32 Mt) and Italy (i.e., 18 Mt), are expected to gain the most in absolute financial terms, with their forecasted savings at the baseline scenario reaching EUR 2.7 million and EUR 1.6 million, respectively. Under a more aggressive, high-price carbon policy, these potential savings increase substantially to EUR 3.9 million for Russia and EUR 2.2 million for Italy, demonstrating that the economic incentive to implement the proposed framework’s recommendations increases proportionally with the strictness of the carbon pricing scheme. The additional financial benefits of even the smaller agricultural sectors, such as Armenia and Slovenia, demonstrate the framework’s scalability, confirming its significance not just for major emitters but also for smaller countries that want to optimize their agricultural footprint in a cost-effective way.

To demonstrate the practical application of the proposed FinTech-aligned optimization framework, this section presents a case study focused on developing targeted decarbonization strategies for the top European GHG emitters. This analysis moves beyond simple emission reduction forecasts to identify the most efficient abatement recommendations for a standardized 5% GHG reduction target. For this analysis, “abatement cost” is defined as the magnitude of the required strategic shift, quantified as the squared Euclidean distance between the baseline and optimized PC scores, serving as a proxy for the level of disruption needed to achieve the target [69]. The “primary abatement strategy” is identified by determining which principal component requires the largest change, thus revealing the key mitigation strategy within each country’s specific profile.

The results, summarized in

Table 5. Optimal decarbonization strategies from the proposed GHG emission optimization framework for each top European GHG emitter.

Country	Emission Profile	Abatement Cost	Abatement Strategy
France	High-Emission Livestock	0.0289	Improve Feed Efficiency
Russia	High-Emission Livestock	0.0166	Improve Feed Efficiency
United Kingdom	Low-Input Traditional	0.0118	Optimize Land Use
Spain	Intensive Cropping and Manure	0.0096	Enhance Fertilizer Use Efficiency
Germany	CH4-Heavy Cropping	0.0087	Re-Balance Agricultural Practices
Italy	Intensive Cropping and Manure	0.0066	Enhance Fertilizer Use Efficiency
Poland	Low-Input Traditional	0.0034	Promote Energy Efficiency
Netherlands	Low-Input Traditional	0.0030	Promote Energy Efficiency
Ireland	Low-Input Traditional	0.0022	Promote Energy Efficiency
Belarus	Low-Input Traditional	0.0012	Promote Energy Efficiency
Romania	Low-Input Traditional	0.0011	Promote Energy Efficiency
Belgium	Low-Input Traditional	0.0004	Promote Energy Efficiency

, reveal that the optimal decarbonization strategy is highly dependent on a country’s existing agricultural profile, validating the value of a tailored, data-driven approach. For countries classified within the “High-Emission Livestock” profile, such as France and Russia, the framework consistently identifies an abatement strategy focused on improving feed efficiency, corresponding to a targeted reduction in the PC1 score (i.e., “Agricultural Production and Emission Intensity”). The mitigation strategy, therefore, involves a direct reduction in methane emissions per unit of livestock product by altering the biological process of enteric fermentation, targeting feed quality enhancement through optimized rations and higher-quality food sources. Interestingly, the abatement cost for France (i.e., 0.0289) is significantly higher than for Russia’s (i.e., 0.0166), suggesting that France’s current system is operating closer to its optimal state within the defined constraints, and achieving the same relative reduction requires a more substantial strategic adjustment.

In contrast, countries with an “Intensive Cropping and Manure” profile, such as Spain and Italy, are advised to enhance fertilizer use efficiency, directly addressing the primary drivers of their emissions by targeting PCs with high positive loadings on nitrous oxide emissions, specifically from crop residue decomposition (i.e., Crops_Residues_N2O) and manure applied to agricultural soils (i.e., Manure_N2O). For these countries, the most effective mitigation strategy involves re-balancing their nutrient management practices, which is achieved through technologies like precision agriculture for variable-rate fertilizer application and the use of slow-release fertilizers or nitrification inhibitors. Italy, in particular, shows a very low abatement cost (i.e., 0.0066), indicating that significant gains can be achieved with relatively minor strategic shifts.

Countries classified under the “Low-Input Traditional” profile showcase an interesting pattern where for the Netherlands, Ireland, Romania, and Belgium that use less-intensive systems, the primary strategy is to optimize energy consumption rather than making drastic changes to land-use or livestock management, while the United Kingdom is suggested to optimize land-use instead. Promoting energy efficiency means reducing the PC3 score (i.e., “Energy Systems Emissions”), which is heavily influenced by variables representing direct carbon dioxide emissions from on-farm energy consumption (i.e., Energy_CO2) and related methane emissions (i.e., Energy_CH4). To achieve this, mitigation strategies could include replacement of diesel farming machinery with electric alternatives, installation of renewable energy sources, or upgrades to high-efficiency irrigation and ventilation systems, depending on specific national characteristics. Finally, Germany, with its unique “CH₄-Heavy Cropping” profile, is assigned a strategy of re-balancing agricultural practices, reflecting the need to address its distinct emission signature, which is heavily influenced by methane from energy systems rather than livestock, and shift its operational focus to reduce this specific emission source.

Based on these results, it is evident that a one-size-fits-all approach to agricultural decarbonization is ineffective. The framework’s main contribution is its ability to demonstrate that the optimal abatement strategy is largely dependent on a country’s pre-existing agricultural profile, recommending feed efficiency improvements for livestock-oriented countries, while identifying energy optimization as the key strategy for lower-intensity systems. However, translating these high-level, data-driven recommendations into widespread, on-farm action presents significant adoption challenges. The framework successfully identifies the strategic “what” (e.g., “Enhance Fertilizer Use Efficiency”), but does not specify the exact operational “how”, which requires extensive agronomic knowledge and local context [70].

These context-dependent outcomes are inherently linked to the framework’s methodological assumptions, particularly the ±20% abatement potential constraint, which serves as a proxy for the proposed structural changes. However, to test the model’s robustness and better understand its behavior under different flexibility scenarios, a sensitivity analysis on abatement potential was also performed, using a more restrictive (i.e., ±10%) [71] and a more expansive (i.e., ±30%) [72] level than our primary selection, as seen in

Table 6. Sensitivity analysis of optimization results for the top 10 European countries with the highest GHG emission reduction potential for 2025 under varying abatement potential constraints.

Country	Emission Reduction (Mt)			Cost Reduction (%)
	±10%	±20%	±30%	±10%	±20%	±30%
Russia	0.00	32.12	0.00	0.00	43.55	0.00
Italy	9.14	18.28	27.41	20.67	41.33	62.00
Spain	6.27	12.55	18.82	11.71	23.42	35.13
Netherlands	4.94	9.88	14.82	15.78	31.56	47.34
United Kingdom	0.00	8.26	12.39	0.00	13.43	20.15
Andorra	2.03	4.07	0.00	1152.75	2305.50	0.00
Cyprus	1.98	3.96	0.00	252.11	504.23	0.00
Luxembourg	1.97	3.93	0.00	217.18	434.36	0.00
Armenia	1.86	3.72	0.00	92.68	185.35	0.00
Slovenia	1.86	3.71	0.00	92.97	185.94	0.00

. For agricultural profiles with significant inherent inefficiencies, such as Italy and Spain, the abatement potential scales directly with the model’s flexibility, suggesting that deeper decarbonization opportunities exist if more long-term structural changes are feasible. On the other hand, the results for Russia and the United Kingdom showcase a threshold effect, where at the more restrictive ±10% level, no improvements are identified, potentially implying that their forecasted trajectories are already optimal within a narrow operational window and meaningful abatement requires larger-scale strategic adjustments. The table also highlights the framework’s limitations, including the difficulty of linear optimization for some countries at the more expansive ±10% level, and the mathematically extreme cost reduction percentages for low-emission countries, such as Andorra and Cyprus. For these profiles, absolute emission reduction is a more stable impact indicator.

The successful implementation of these strategies ultimately depends on overcoming socio-economic barriers, including the direct financial implications for farmers, such as the significant upfront investment costs for precision agriculture technologies or higher-quality feed sources. Overcoming these challenges requires not only access to knowledge and training [9], but also the presence of supportive agricultural policies. Therefore, while the proposed framework provides a powerful, data-driven starting point for prioritizing mitigation efforts, its practical impact depends on integrating these strategic insights into targeted policy design, effective financial incentives, and extension services that can maintain a robust IoT infrastructure in rural environments and, subsequently, guide its implementation [73].

7. Conclusions and Future Work

This study proposed an explanation-focused GHG mitigation framework for IoT-enabled agriculture by implementing a two-stage methodology combining statistical modeling, using PCA and k-Means clustering, with a cluster-oriented PCR model and a constrained optimization routine. The GHG emission drivers analysis successfully identified four distinct emission profiles within the European context, each with a unique emission signature. The experimental results showcased that the cluster-oriented PCR model significantly outperformed its generalized variant, with accuracy improvements of up to 55.7%, confirming that a profile-specific approach provides reliable emission modeling. The subsequent FinTech-aligned optimization revealed substantial, context-dependent abatement potential, with forecasted cost reductions as high as 43.55% for certain profiles, while also identifying that some countries were already operating near their short-term optimum. The case study further illustrated the framework’s ability to translate these numerical outputs into specific, interpretable strategies, providing a transparent, auditable, and economically grounded pathway for generating strategic decarbonization recommendations that are directly linked to the underlying drivers of emissions.

The proposed optimization framework could be further refined and validated with the integration of more granular, farm-level (Tier 2/3) data sourced from real-world IoT-enabled agricultural testbeds. In addition, while the linear nature of PCR was well-suited for the Tier 1 dataset, exploring non-linear regression techniques for the more complex clusters could provide further accuracy improvements. These future extensions will also focus on translating the framework’s insights into operational guidance at the farm level, integrating the framework with real-world FinTech platforms for green financing.