Carbon Emission Intensity Modeling in Freight Operations Using eXplainable AI in Real-World Conditions

Abstract

The present study evaluates the effect of real-world operational factors and driving behaviors that significantly contribute to CO₂ emissions and total energy consumption of the port-based heavy-duty vehicles (HDVs). Interpretable machine learning techniques are applied within an eXplainable Artificial Intelligence (XAI) framework to assess the impact of input variables on prediction accuracy. The inherent simplifications in these approaches often limit their ability to capture the complex, nonlinear characteristics of vehicular emission determinants, particularly under dynamic, micro-operational conditions associated with real-world settings. XGBoost showed higher predictive accuracy over conventional regression and other ensemble methods, with up to 46% improvement in R ² and over 80% reduction in estimation errors. To address the black-box nature associated with the model, this study adopts XAI techniques, with SHapley Additive exPlanations (SHAP) employed to quantify feature contributions and enhance the interpretability. The results show that real-world CO₂ emission levels remain substantially high under dynamic operational conditions, emphasizing the need for improved transit and freight management strategies to mitigate vehicular emissions. This further reinforces the importance of regulatory frameworks that incorporate CO₂ emission and fuel-efficiency standards alongside conventional pollutant limits. Such progressive targets are intended to curb the climate impact, stimulate technological innovation, and support long-term low-carbon transition goals.

1. Introduction

The transport sector is a leading energy consumer and remains a key source of air pollutant emissions. Globally, heavy-duty vehicles constitute a limited portion of the vehicle fleet, yet account for a substantial share of fuel consumption and greenhouse gas emissions due to their energy-intensive operations. This sector accounts about 70% of the nation’s diesel consumption. From the 1950s to 2011, road transport in India experienced substantial growth, with the passenger share increasing from 15 to 86% and the freight share from 14 to 70%. HDVs serve as an essential connection between long-haul freight movement and last-mile distribution, which improves overall operational efficiency. Recent studies have examined the contribution of HDVs to environmental degradation, especially in urban and industrial transport corridors. Globally, diesel consumption among HDVs varies substantially across regions despite similar fleet sizes, fuel intensity varies due to differences in freight activity, operational patterns, and vehicle technologies. Heavy-duty vehicles represent a major contributor to global fuel demand, with consumption patterns varying widely between countries due to differences in freight intensity and operational activity. Figure illustrates the 2020 HDVs fuel consumption estimates, based on the IIASA GAINS model and standardized using a diesel energy content of 35.8 MJ/L, emphasize the regional disparities in fuel use. , Heavy-duty trucks account for approximately 20% of the total fuel consumption in the United States, underscoring their disproportionately high energy demand in freight transport. In the European Union, heavy-duty vehicles contribute over 25% of road transport-related greenhouse gas (GHG) emissions and account for more than 6% of the EU’s total GHG output. While they are vital to supply chain operations, their emissions were often concentrated near freight corridors, disproportionately affecting communities located near ports, warehouses, and major roads. The steady rise in cargo traffic at Indian ports (Supplementary Figure S1), indicates a sustained intensification of freight activity in recent years, driven by increased demand for bulk commodities and containerized goods. Residents of Low and Middle-Income Countries (LMICs) were often located near major roads and freight hubs, resulting in greater exposure to pollutants from heavy-duty vehicles and increased vulnerability to air-quality-related health risks. Studies in port settings identified the concentrated environmental impacts of HDV activities and suggested targeted operational controls, multimodal solutions, and fleet optimization to lower CO₂ emissions. , Heavy diesel trucks used in freight operations emit about 2.6 kg of CO₂ per liter of fuel burned. The escalating fuel demand underscores the urgent need for sustainable transport solutions and energy policies. , The European Parliament and the Council introduce mandatory CO₂ emission performance standards for new heavy-duty vehicles, including a required 15% reduction in average fleet emissions by 2025 (relative to the 2019 baseline) and more ambitious targets by 2030/2035. These targets are designed to mitigate the climate impact of HDVs, stimulate technological innovation, and align the transport sector with the long-term low-carbon transition goals.

1.

Global diesel consumption by heavy-duty vehicles (HDVs) in 2020, represented country-wise. The total fuel consumption (in million liters), highlighting significant regional differences in freight-related energy use.

Previous studies showed that machine learning (ML) techniques improved the accuracy of carbon emission forecasting and supported policy decisions through greater model transparency and interpretability. , However, challenges such as inconsistent data, limited model clarity, and lack of standardized evaluation remained significant limitations. Table presents a summary of the methodological approaches used to model and forecast CO₂ emissions in the transportation sector. It was observed that micro-operational trends under real-world conditions had received limited attention in previous studies, leaving useful gap in understanding emissions more precisely. Many studies have successfully modeled transport sector GHG emissions at the macro level using diverse methods and annual predictor data sets such as GDP, population, energy use, and vehicle manufacturing data. − These approaches have provided valuable insights into long-term emission trends, though their reliance on broader data sets highlights opportunities to enhance temporal granularity and capture real-world heterogeneity for localized or dynamic scenarios. Similarly, several recent studies employed national inventory data sets derived from manufacturer-reported laboratory driving cycles to estimate fuel consumption and CO₂ emissions from light-duty vehicles. − While these data sets offer important standardized benchmarks, their representation could be further strengthened by integrating higher temporal resolution and real-world operational dynamics to better reflect localized and dynamic transport conditions.

1. Review of Previous Literature on Modeling Transport-Based Energy and Carbon Emissions.

study	time span	region	evaluation metrics	temporal granularity	data sets/variables	model	model strengths	model limitations
	2005–2015	Taiwan	MAPE	annual	GDP, urban population, energy consumption	MGPM	interpretable, suitable for long-term trends	nonlinear capability
	1970–2016	Türkey	R 2, MAE, MAPE, MSE, RMSE, MBE	annual	energy consumption, population, GDP, year, vehicle kilometer	XGBoost	high accuracy, handles nonlinear relationships	interpretability, tuning-sensitive
	1990–2017	China	APE, MAPE	annual	population, economy, energy, urbanization, transport, industrial structure	PSO+GM	effective for limited data sets, trend forecasting	complex nonlinear dynamics
	2020–2023	UK, Canada	MAE, MAPE	annual	engine power, fuel consumption, CO2 emission, engine capacity	DTR	interpretable decision structure	prone to overfitting, unstable
	1995–2014	Canada	R 2, MSE, RMSE, MAPE	annual (7 yr compiled)	engine size, fuel consumption, cylinders, CO2 emission, make, model, fuel type	MLP-SHAP	captures nonlinearity, interpretable	data-intensive, computationally demanding
	1995–2014	Canada	R 2, MSE, RMSE	annual (7 year compiled)	engine size, fuel consumption, cylinders, CO2 emission, make, model, fuel type	BiLSTM	temporal dependency modeling	black-box, high training requirement
	1997–1999	India	RMSE, MBE, MSE, R 2	hourly	vehicular source strength (CO & NO2), meteorology–17 feature variable	ANN	flexible nonlinear modeling	transparency, overfitting risk
	2014	Poland	MSE, R 2	per second	CO2, velocity, acceleration, engine load, RPM, altitude, manifold, turbo boost	XGBoost	robust for high-resolution data	feature engineering sensitive
	2014–2020	Canada	R 2, RMSE, adj. R 2	annual	engine size, cylinders, fuel type, consumption, transmission, make, class, CO2	XGB, RF, LGBM	ensemble robustness, strong predictive power	reduced physical interpretability
	2011–2013	Japan	R 2, MAE, RMSE	per minute	energy consumption, trip distance, speed, air conditioner, ambient temperature, road gradient	LightGBM	fast training, handles large data sets	weaker temporal memory
	2020	Canada	R 2, RMSE	per minute	speed, density, traffic, flow, delay, CO2	LSTM	effective for sequential traffic patterns	data-hungry, limited explainability
	2015–2016	India	RMSE, MBE, IA	per second, 5 h	vehicle speed, acceleration, VSP, engine RPM, oil temperature	NARX	captures dynamic system feedback	sensitive to noise, parameter selection
	2010–2013	US	CV, MAD, SD	30 s	speed, accel. /decel., driving events and behaviors	K-means, Autonomie	behavior pattern discovery	nonpredictive, cluster-dependent
	1990–2019	UK	RMSE, rRMSE,MAPE, MAE	annual	energy consumption, population, electricity, unemployment and other 24 features	SVR-RBF, RF, SHAP	balanced accuracy and interpretability	computational complexity
	1970–2021	Turkey	MAE, RMSE, R 2, sMAPE	annual	CO2, Oil consumption, GDP, population, electricity, vehicle count, energy supply	ANN, GBDT	strong nonlinear learning	reduced physical interpretability
	2024	China	R 2, RMSE	5 s	speed, VSP, acceleration, road slope, CO2	DL-DTCEM	high-resolution dynamic modeling	limited generalizability, high data demand

Advanced freight management relies on improving operational efficiency within a sustainable framework. Yet, limited real-world evidence on HDV operations restricts policy strategies to curb energy and emission footprints. Carbon emissions from road transport were forecasted using freight activity and macro-modal trends, while the uncertainties associated with long-term projections were acknowledged due to changing economic and policy dynamics. , An Artificial Neural Network (ANN)-based line source model was introduced to predict urban vehicular emissions using field data on traffic flow, vehicle classification, and meteorological parameters, highlighting the effectiveness of machine learning in analyzing large-scale and complex emission data sets. Gradient boosting regression was applied to assess transportation-related CO₂ emissions at the macro level, identifying road density, industrial activity, and vehicle ownership as the primary contributing factors. Carbon emissions from road freight transport were demonstrated that tree-based models, particularly decision trees, outperformed traditional linear approaches in predicting CO₂ emissions from vehicles, offering improved accuracy in emission estimation. A hybrid approach integrating machine learning techniques to forecast urban carbon emissions was proposed and demonstrated approximately 20% improvement in predictive accuracy over individual models, while data sparsity in some regions posed challenges. GPS-based microtrip analysis with XGBoost was employed, trained on Motor Vehicle Emission Simulator (MOVES) data to predict emissions. A novel carbon emission forecasting method that integrates the environmental kuznets curve hypothesis with a nonlinear multivariate gray model was developed to improve prediction accuracy for transportation sector emissions. However, the generalizability of the approach was limited by variations in operational factors and applicability across other regions. An ensemble model integrates Random Forest (RF) and XGBoost for transient CO₂ and NO _x prediction, identifying real-time driving variables as key emission determinants. The application of machine learning, particularly XGBoost, in microscale dynamics improved emission prediction accuracy compared to traditional models. , Model-agnostic explanations tool like SHAP, and partial dependence plots, while facilitating evidence-based policy development with their visual Interpretability enabling effective communication of complex model outputs. −

The limited interpretability in machine learning models poses significant challenges, especially in contexts requiring transparent and explainable decision making. While SHAP had recently been applied across various domains. , This study develops an integrated, data-driven framework to link real-world driving characteristics with CO₂ emission profiles, using micro-operational data collected from 14 HDVs engaged in port-based cargo transfer. By addressing the limitations in capturing complex, nonlinear emission dynamics, the framework applies interpretable machine learning techniques within an eXplainable AI (XAI) setting to evaluate variable impacts on predictive performance. This approach proposes an understanding of emission-sensitive operational regimes, offering an analytical basis for targeted freight management strategies, operational efficiency, and measures to curb emissions from HDV fleets.

2. Materials and Methods

2.1. Study Site and Measurement Design

Field measurements were performed at Chennai port, one of the largest and oldest major ports in India, situated along the Bay of Bengal, as shown in Figure . Covering around 590 acres of land and 420 acres of water, the port functions as a major hub for container trade, automobile exports, and bulk cargo along eastern coast of India. It is equipped with three docks, 24 berths, and a draft depth ranging from 12 to 16.5 m, supporting an annual cargo handling capacity of up to 60 million tonnes.

2.

Geographic location and layout of Chennai Port, India. Field photographs showing port operations, test vehicles, and on-board measurement systems were taken by the authors. The Chennai Port layout map was sourced from the Chennai Port Authority Web site.

Measurements under real-world vehicle operating conditions were conducted to evaluate effectiveness specific to port activities. The monitoring campaign took place during routine daytime operations (10:00–16:00) at the port terminal, aligning with scheduled cargo handling activities. Each monitored trip included multiple loops along a predefined internal route, encompassing both loading and unloading operations. Vehicle movement along the berths was organized as a two-way traffic system. The study involved 14 heavy-duty diesel vehicles (HDDVs) comprising tractor–trailer configurations broadly corresponding to Class 7–8 vehicles, as summarized in Table . The container used in port operations had a maximum gross weight of 30.48 tonnes, including a tare weight of approximately 2.19 tonnes. In routine operations, these vehicles typically carry payloads ranging from approximately 18 to 30 tonnes, depending on container size, cargo density, and operational constraints. Fully loaded operations generally involve a single 20-ft container, corresponding to a standard ISO container, commonly referred to as a Twenty-foot Equivalent Unit (TEU), as summarized in Supplementary Table S1. The TEU is a standardized unit of measurement used in shipping and port operations, whereas off-loaded operations involve empty containers or trailers with negligible payload (without containers). The reported vehicle weight range reflects real-world port-operational conditions, where containers are frequently moved in partially loaded or intermediate logistics stages and are constrained by axle-load compliance and safety considerations. The monitored vehicles were selected to represent the dominant vehicle configurations used in routine port operations. These internal transfer vehicles (ITVs) include both articulated (tractor-trailer) and rigid vehicles employed for container and freight movement within the port terminal. Key parameters such as vehicle age (in years), cumulative mileage (in kilometres), and curb weight (in tonnes) were manually recorded. Vehicles were not fitted with any advanced after-treatment systems. A total of 375,000 s-by-second data samples were collected. GPS and emissions data were recorded independently and later synchronized to align timestamps, enabling analysis of vehicle position, activity patterns, and associated emissions. All field monitoring adhered to standardized safety protocols in accordance with the International Maritime Dangerous Goods (IMDG) code, issued by the International Maritime Organization, ensuring safe practices during container and commercial cargo handling.

2. Summary of Internal Transfer Vehicle Characteristics Monitored during Port Operations .

vehicle ID/ITV	engine disp. (cc)	model year	emission std.	cabin	unladen weight (kg)	chassis type	load state
1	5660	8 yr 10 m	BS-III	dual cab	15,320	articulated	unloaded
2	5883	6 yr 10 m	BS-IV	dual cab	15,560	rigid	unloaded
3	5700	7 yr 4 m	BS-IV	dual cab	15,560	rigid	unloaded
4	5900	7 yr 8 m	BS-IV	single cab	13,900	articulated	unloaded
5	5883	6 yr 10 m	BS-IV	dual cab	14,740	articulated	unloaded
6	5660	3 yr 6 m	BS-IV	dual cab	15,320	articulated	unloaded
7	5660	7 yr 11 m	BS-III	single cab	13,900	articulated	unloaded
8	5700	5 yr 3 m	BS-IV	dual cab	15,560	rigid	loaded
9	5660	9 yr 10 m	BS-III	dual cab	15,320	articulated	loaded
10	5660	8 yr 9 m	BS-III	dual cab	15,560	rigid	loaded
11	5700	9 yr 10 m	BS-III	single cab	14,020	rigid	loaded
12	5660	7 yr 8 m	BS-IV	dual cab	15,560	rigid	loaded
13	5660	6 yr 2 m	BS-IV	single cab	14,020	rigid	loaded
14	5990	10 yr 4 m	BS-III	dual cab	15,560	rigid	loaded

^aAbbreviations: yr = years; m = months; cc = cubic centimeters. Articulated refers to HDVs with a jointed chassis (tractor–trailer configuration).

2.2. On-Board Emission and Fuel Use Characterization

Vehicular emission characteristics were derived from second-by-second operational parameters. Vehicle speed and altitude data were concurrently recorded at 1 Hz using hand-held GPS units (Garmin e-Trex 20X). To ensure accuracy, on-board measurement, acceleration–deceleration (AD) and positional data were time-synchronized and logged at a one-second resolution. Emission measurements were performed using an on-board portable emissions measurement system (PEMS), which consisted of an AVL Ditest, a certified exhaust gas analyzer approved by Automotive Research Association of India (ARAI) for testing and validation. An exhaust gas sampling probe, control unit, and data logger were integrated for real-time acquisition and powered by an uninterruptible power supply (UPS). The components were installed in the driver compartment (Figure ). It utilizes an infrared-based NDIR technique for CO₂ detection. The system was calibrated for zero and span both before and after each test cycle, resolution with an accuracy range of ±0.3% vol for CO₂. The vehicular emission characteristics were determined from instantaneous operational metrics, as given in eq .

$$E(\theta )=f\left(V_x\right)$$ 1

where, E(θ) represents the emission factor for pollutant (θ), while V _x denotes the vehicle operational parameters such as speed, exhaust flow, air-fuel ratio (AFR). The emission factor was determined from instantaneous CO₂ concentration measured at one-second intervals, expressed in either volume percentage (% vol) or parts per million (ppm).

Fuel consumption was estimated with the carbon balance method (eq ), based on the emission rates of carbon pollutants. The adopted carbon mass fractions were 0.273 for CO₂, 0.429 for CO, and 0.866 for HC.

$$\mathrm{FR}(t)=[0.273\times {\mathrm{ER}}_{{\mathrm{CO}}_2}(t)+0.429\times {\mathrm{ER}}_{\mathrm{CO}}(t)+0.866\times {\mathrm{ER}}_{\mathrm{HC}}(t)]/W_{\mathrm{C}}$$ 2

$${\mathrm{FE}}_{\mathrm{bulk}}=\frac{360\times \mathrm{FR}}{{\rho }_{\mathrm{d}}\times \nu }$$ 3

where, FR is the average fuel consumption rate (g/s) for vehicle, t in seconds, and W _C is the fuel carbon content coefficient (0.866 for diesel). Furthermore, distance-specific fuel efficiency for vehicles was calculated in liters per 100 km using eq , for diesel density ρ_d = 0.85 kg/L, following USEPA. FE_bulk is the bulk fuel economy in L/100 km, and ν in km/h.

2.3. Machine Learning Techniques

Description of the applied ML techniques used in the analysis is presented in the subsequent section. The hyperparameter search space and selected optimal values for each model are summarized in Table .

3. Search Space for Hyperparameter Tuning.

model	search space
MLR	linear: ordinary least-squares ∈ no regularization
LASSO	alpha: L1 penalty strength; controls sparsity ∈ 0.0012
Bayesian Ridge	alpha_1: prior over weights (gamma distribution) ∈ 1 × 10–6
	alpha_2: prior over noise (gamma) ∈ 1 × 10–4
	lambda_1: prior over weights precision ∈ 1 × 10–4
	lambda_2: prior over noise precision ∈ 1 × 10–6
Random Forest	learning rate (η): [0.01–0.3] ∈ 0.07
	N estimators (Trees): [50–300] ∈ 200
	maximum tree depth: [3–20] ∈ 19
	min samples split: [0.5–1.0] ∈ 6
	min samples leaf: [0.5–1.0] ∈ 1
	max features: [auto, sqrt, log2] ∈ sqrt
XGBoost	learning rate (η): [0.01–0.3] ∈ 0.07
	N boosting rounds (trees): [50–300] ∈ 250
	gamma: [0–0.5] ∈ 0.048
	maximum tree depth: [3–10] ∈ 7
	subsample ratio of training instances: [0.5–1.0] ∈ 0.6
	subsample ratio of columns by tree: [0.5–1.0] ∈ 0.8

2.3.1. MLR

Multiple Linear Regression (MLR) is a standard technique for predicting the value of a dependent variable Y _i from a set of independent variables X _ij . The model is shown in eq .

$$Y_i=\sum _{j=1}^k{\beta }_j{X}_{ij}+{\epsilon }_i$$ 4

where, k is the total number of independent variables, β _j indicates the coefficient for variable j, X _ij is the value of the jth variable for observation i, and ε _i is the residual error term.

2.3.2. Bayesian Ridge

Bayesian ridge regression (BRR) estimates probability distributions for the parameters, taking into account data uncertainty and using prior knowledge to make the estimates more reliable is computed from eq .

$$\beta =\sum _{i=1}^n{[y_i-f\left(x_i\right)]}^2+\lambda \sum _{i=1}^n{{\beta }_i}^2$$ 5

where, f(x _i ) represents the model prediction for the ith observation, incorporating prior knowledge about the parameters before observing the data. The term $\sum _{i=1}^n{(y_i-f\left(x_i\right))}^2$ denotes the total residual sum of squares (RSS) across all n observations and $\sum _{i=1}^n{\beta }_i^2$ , is the sum of squared regression coefficients, which, together with the RSS, balances data fit and the L2 regularization penalty.

2.3.3. LASSO

In the Least Absolute Shrinkage and Selection Operator (LASSO) regression, the ridge penalty is replaced with an L1 norm penalty defined by $\lambda \sum _{i=1}^M\left|{\beta }_{h,i}\right|$ . This gives the estimator in eq used to calculate the model’s coefficients.

$$\widehat{{\beta }_h}={\mathrm{argmin}}_{{\beta }_h}\left[\sum _{t=1}^{T-h}(y_{t+h}-{\acute{\beta }}_h{x}_t{)}^2+\lambda \sum _{i=1}^M|{\beta }_{h,i}|\right]$$ 6

The model selects features whose coefficients remain nonzero after regularization. Notably, LASSO can perform effectively even when the number of features exceeds the number of samples. Here, T denotes the sample size, M is the total number of features, and λ is a hyperparameter that controls the degree of shrinkage.

2.3.4. Random Forest

Random Forest (RF) is an ensemble method that builds multiple decision trees and combines their results. It uses information gain or impurity to determine the best split at each node. In decision trees, nodes contributing the highest impurity reduction typically appear near the root, whereas those with lower reductions are found toward the leaves. Consequently, selective pruning at specific nodes can help extract a subset of the most relevant features.

2.3.5. XGBoost

Extreme Gradient Boosting (XGB) employs second-order (newton) tree boosting to fit additive decision tree models. In this framework, decision trees act as base learners, performing binary feature splits (if-else conditions) to either classify samples in classification tasks or predict continuous outcomes in regression tasks. This approach provided evaluation of generalization performance, as shown in Figure . For a data set containing n examples, D = {(x _i , y _i )} _i = 1 , the prediction from a single regression tree f _t at iteration t (eq ).

$$f_t\left(x_i\right)=\sum _{j=1}^T{w}_{j}I(x_i\in R_j)$$ 7

where R _j denotes the region (leaf) j, w _j is the leaf output, and I is the indicator function. XGB optimizes its learning process by incrementally adding f _t (x _i ) to the ensemble of base learners to minimize the regularized objective function, whose second-order approximation at iteration t (eq ).

$${Ũ}^{(t)}=\sum _D\sum _{j=1}^T\left[g_i{w}_j+\frac{1}{2}{h}_i{w_j}^2\right]=\sum _{j=1}^T\left[G_j{w}_j+\frac{1}{2}{H}_j{w_j}^2\right]$$ 8

where $G_i=\sum _D{g}_i$ , $H_i=\sum _D{h}_i$ , and g _i and h _i are the first and second derivatives of the loss function w.r.t. the prediction. In each iteration, the optimal leaf weights w _j for a fixed tree structure are derived by setting the derivative of the objective to zero (eq ).

$$w_j^{*}=-\frac{G_j}{H_j},j=1,...,T$$ 9

4.

Model evaluation of the XGBoost for CO₂ emissions (kg hr^–1). (a) Actual vs predicted for the training set (left) and test set (right). (b) Residual plots for using both training (5-fold cross-validation) and test data sets.

Building the tree structure, XGB evaluates candidate splits based on their ability to reduce the objective (eq ). The optimal split is determined by calculating the total score of a tree, The gain from a potential split is computed as eq .

$${\mathrm{obj}}^{*}=-\frac{1}{2}\sum _{j=1}^T\frac{{G_j}^2}{H_j}$$ 10

$$\mathrm{Gain}=\frac{1}{2}\left[\frac{{G_{\mathrm{L}}}^2}{H_{\mathrm{L}}}+\frac{{G_{\mathrm{R}}}^2}{H_{\mathrm{R}}}-\frac{G^2}{H}\right]$$ 11

where, G _L, H _L and G _R, H _R correspond to the left and right child nodes, respectively. Splits that result in negative or low gain are pruned bottom-up to avoid overfitting and reduce complexity.

2.4. Analytical Framework of Integrated Model Development

This section outlines the analytical procedures used to evaluate the performance of the machine learning models, including feature extraction, training-validation, and accuracy assessment through statistical and error-based metrics. Analyses were performed on instantaneous fuel consumption, real-world operational parameters, and corresponding CO₂ emissions. Normal probability and Q–Q plots were used to evaluate the distributional characteristics of the features and to determine the extent to which they conformed to normality (Figure S3). The kolmogorov–smirnov test reported a statistic of 0.127 (p-value < 0.05) across speed bins, confirms a significant deviation of the CO₂ emission data from normality (Figure S5), thereby justifying the use of nonlinear modeling approaches. A two-factor ANOVA was performed to evaluate the influence of speed bins and vehicle loading states on CO₂ emissions and fuel consumption using a large micro-operational data set (Table S4). The resulting p-values for all comparisons were below 0.05, indicating statistically significant differences in mean emissions and fuel use across speed bins and between loading conditions. Pearson correlation coefficients were calculated to identify potential multicollinearity among independent variables (Figure ). Finally, all numerical variables were scaled to a standard range using min-max normalization. While the focus of this study is on predicting CO₂ emissions under real-world operational conditions, it also aims to apply a multilevel data approach integrated with ML to develop and compare five models. Comparative modeling was conducted using MLR, LR, BRR, RF, and XGB. The algorithms were benchmarked to evaluate predictive performance in terms of accuracy, reliability, and generalization, facilitating the identification of key influencing predictors for optimal model selection. , In this approach, tree models act as base learners, applying if-else or true-false feature splits to classify data points in classification problems or predict continuous values in regression problems. The algorithmic architecture, as illustrated in Figure , depicts the input feature vectors (x) and the independent and identically distributed random vectors (w _j ). The XGBoost model aggregates the weighted contributions of all individual decision trees to generate the final predictive output. Furthermore, partial dependence and SHAP analysis was employed to assess the impact of individual features on model performance, grounded in game theory principles, this approach assigns each variable a weighted contribution, reflecting in the model predictions.

3.

Correlation matrix (pearson) of input features.

5.

Schematic representation of the ML based modeling framework.

2.5. Hyperparameter Settings

Effective hyperparameter optimization is critical for improving the performance of machine learning models. A preliminary study was conducted to manually optimize parameters in order to assess model performance and analyze the influence of different features on CO₂ and energy consumption. The optimal hyperparameter values identified during model tuning for each ML model are presented in Table .

2.6. Model Evaluation Metrics

In this study, subsequent data processing and computational analysis were carried out using Google Colab, a cloud-based open-source python environment that supports jupyter notebooks, for the training and testing of model. To evaluate model performance, 5-fold cross-validation was employed by partitioning the data set into five equal subsets. The model was iteratively trained on four folds and tested on the remaining fold, rotating until each subset served as the test set once. The model selection process involved evaluating all possible combinations of predictor variables, with the optimal model determined based on four key performance metrics, the coefficient of determination (R ²), root mean squared error (RMSE), mean absolute error (MAE), and mean squared error (MSE), as defined in eqs and .

$$R^2=1-\frac{\sum _{i=1}^n[y(u_i,v_i)-ŷ(u_i,v_i){]}^2}{\sum _{i=1}^n[y(u_i,v_i)-\mathrm{y̅}(u_i,v_i){]}^2}$$ 12

$$\mathrm{RMSE}=\sqrt{\frac{1}{n}\sum _{i=1}^n[y(u_i,v_i)-ŷ(u_i,v_i){]}^2}$$ 13

$$\mathrm{MAE}=\frac{1}{n}\sum _{i=1}^n|[y(u_i,v_i)-ŷ(u_i,v_i)]|$$ 14

where n is the number of observations, y(u _i , v _i ) is the actual value for observation, ŷ(u _i , v _i ) is the predicted value for observation, $\stackrel{\circledR }{y}(u_i,v_i)$ is the mean of the actual values.

2.7. Shapley Additive Explanation (SHAP) Based Feature Interpretation

SHAP integrates optimal credit allocation with localized explanations by applying the Shapley value concept from cooperative game theory. Specifically, SHAP provides individual-level model interpretation for a given prediction, f(x), by decomposing it into the sum of feature contributions and the baseline prediction, as expressed in eq :

$$f(x)={\varphi }_0(f,x)+\sum _{k=1}^K{\varphi }_k(f,x)$$ 15

where, ϕ₀(f, x) represents the expected value (mean) of the model predictions across the entire data set, while ϕ _k (f, x) denotes the Shapley value quantifying the contribution of the kth feature to the specific prediction f(x), relative to the baseline ϕ₀(f, x). The Shapley value method fairly distributes the contributions of explanatory variables by averaging their marginal contributions across all possible feature subsets for each instance. The relative significance of the kth feature, I _k ^mode is obtained by averaging its absolute Shapley values across all n samples computed from eq :

$$I_k^{\mathrm{mode}}=\frac{1}{n}\sum _{i=1}^n\left|{\varphi }_k{(f,x)}^i\right|$$ 16

Subsequently, the feature importance for predicting mode shares is aggregated across all relevant variables and normalized to show relative importance, as indicated in eqs and :

$${\stackrel{-}{I}}_k=\sum _k\frac{1}{9}(I_k^{\mathrm{speed}}+I_k^{\mathrm{accel}./\mathrm{decel}.}+I_k^{\mathrm{fuel}\mathrm{eff}.}+I_k^{\mathrm{mileage}}+I_k^{\mathrm{engine}\mathrm{disp}.}+I_k^{\mathrm{age}}+I_k^{\mathrm{weight}}+I_k^{\mathrm{AFR}}+I_k^{\mathrm{altitude}})$$ 17

$${\mathrm{I̅}}_k=\frac{{\mathrm{I̅}}_k}{\sum _k{\mathrm{I̅}}_k}\times 100\%$$ 18

SHAP dependence matrix interprets the relationship between specific features and their contribution, illustrating their directional impact on the mode share predictions. Each point in the plot corresponds to an observation in the data set, where positive SHAP values indicate a positive contribution of the feature to the predicted mode share (Figure ), whereas negative values imply a suppressive influence.

6.

SHAP summary plot illustrating the relative influence of features on predicted emissions.

3. Results and Discussion

3.1. Effect of Operational Condition on Real-World Emissions

A preliminary statistical analysis was conducted to evaluate key operational features and CO₂ emissions of ITVs under loaded (with container payloads) and off-loaded (empty containers or negligible payload) operating conditions, as presented in Table . Vehicles operating under load exhibited a lower mean speed (7.12 km/h) compared to off-loaded conditions (10.31 km/h), due to more restricted mobility during cargo transport in port settings. The mean acceleration was close to zero in both cases, consistent with frequent idling and stop-and-go behavior. Off-loading operations show a higher average speed and acceleration (Figure S2), with vehicles spending more time in cruise mode (57.31%) and less in idle (23.72%) compared to loading operations presented in Table S3. Conversely, loading conditions were characterized by significantly higher idle time (36.41%) and lower speeds, suggesting more frequent stops and slower manoeuvring during cargo handling. Fuel consumption differs notably under load and off-load conditions, as shown in Figure S4. The load scenario exhibited a longer right tail (>10 L/h) in the Gaussian kernel density estimation, suggesting periods of higher, less efficient fuel usage. Fuel consumption rates were markedly higher under load than off-load. Similarly, CO₂ emissions under load averaged 12.79 kg/h was approximately twice that recorded in off-load conditions. The vehicle fleet operating under load conditions was older and had higher cumulative mileage than those in off-loaded operation, potentially contributing to higher fuel consumption and emissions. Although both conditions involved engines of similar displacement, AFR with minor skewness variations suggesting subtle differences in combustion dynamics. Changes in altitude were minimal and symmetrically distributed over the level terrain of port corridors.

4. Descriptive Statistics of Operational and Vehicular Attributes in Real-World Conditions.

(a) Load Condition
N obs. (14 ITV)	mean	std dev.	min	0.25	0.75	max	variance	skewness
speed (km hr–1)	7.12	6.97	0	0	11.99	30.5	48.59	0.63
acceleration (m s–2)	0	0.21	–1.4	0	0.1	2	0.04	–0.63
fuel efficiency (L h–1)	5.32	3.16	0	1.85	6.72	29.74	9.97	0.67
mileage (103 km)	44.57	9.15	37.41	37.9	56.52	62.36	83.61	1.02
engine (cm3)	5719.67	113.89	5600	5600	5900	5900	12.74	0.65
age (months)	98.66	17.86	74	84.3	117	124	319.2	–0.03
AFR	21.54	5.81	14	17.64	25.11	33.6	33.74	0.68
Δaltitude (m)	–0.04	0.72	–4.8	–0.3	0.2	4.8	0.52	0.58
CO2 (kg h–1)	12.79	8.79	0	4.93	18.66	36.71	77.24	0.63

(b) Off-Load Condition
N obs	mean	std dev.	min	0.25	0.75	max	variance	skewness
speed (km hr–1)	10.31	7.33	0	2.82	15.71	34.53	53.77	0.07
acceleration (m s–2)	0	0.12	–1.5	0	0	1.9	0.01	0.43
fuel efficiency (L h–1)	3.61	2.1	0	1.52	4.52	17.29	4.39	1.45
mileage (103 km)	36.75	7.78	23.79	29.07	43.58	47.22	60.85	–0.1
engine (cm3)	5755.29	117.58	5600	5700	5900	5900	13.2	0.2
age (months)	70.54	21.74	42	53	85.2	106	472.6	0.27
AFR	21.8	6.82	14	14	26.56	33.61	46.56	0.15
Δaltitude (m)	–0.01	0.59	–4.9	–0.1	0.1	4.7	0.35	–0.32
CO2 (kg h–1)	8.32	5.61	0	3.98	11.27	34.81	31.47	1.46

3.2. Multicollinearity Analysis

Multicollinearity affects model interpretability and stability by making it difficult to isolate the individual impact of independent variables. It can distort coefficient estimates and reduce the statistical significance of predictors. Table presents the Variance Inflation Factor (VIF) and tolerance for the selected input features. speed, acceleration, AFR, and altitude exhibit low VIFs, indicating minimal multicollinearity. The data set was iteratively pruned by removing highly collinear features in regression models until all variables exhibited VIF values below 10. Fuel efficiency and CO₂ emissions were strongly correlated and show high VIFs (>7), suggesting redundancy (Figure ). Slight trend indicated that larger engines were associated with older vehicles. A moderate link between speed and CO₂ suggests its contribution to emission variability, while mileage, engine size, and age exhibited intercorrelations (r = 0.5–0.7) alongside acceptable VIFs.

5. Multicollinearity Analysis.

feature	speed	accel.	fuel eff.	mileage	engine	age	weight	AFR	ALT	CO2
VIF	1.22	1	8.3	3.84	4.11	4.71	3.62	1.11	1	7.69
tolerance	0.82	1	0.12	0.26	0.24	0.21	0.28	0.9	1	0.13

3.3. Nonlinear Interactions of Predictors in ML

The XGBoost model showed high predictive performance, as indicated by steadily decreasing RMSE with larger training sets and close alignment between training and validation errors (Figure ). Though some dispersion was observed at higher predicted values, the residuals remained homoscedastic. Speed-bin-wise fuel use and CO₂ emission characteristics under real-world load conditions (Table S2). In 0–10 km/h range, emissions remain relatively low, with a mean of 6.98 kg/h, though some occasional spikes were noted, likely due to idling and stop and go scenarios. Emissions increased sharply in the 10–20 km/h range (mean 11.88), corresponds to initial acceleration. In the 20–30 km/h range, mean emissions decreased but with peaks up to 36.71 kg/h, indicating variability. Emissions were highest beyond 30 km/h with a mean of 18.24 kg/h, consistent with elevated fuel demand at sustained speeds. Emissions increased as speed rises (Figure ), with the most pronounced and consistent emissions observed at speeds exceeding 25 km/h

7.

(a) Learning curve of the XGB regressor. (b) Training and testing RMSE progression across epochs.

10.

Spatial interpretation of vehicle speed, fuel efficiency, and partial dependence of predicted CO₂ emissions (kg hr^–1) under (a) off-load and (b) load operations at the port terminal.

Predicted CO₂ emissions show strong correspondence with real-world driving operations. Fuel consumption, speed, load proxies (mileage, age, engine displacement), and acceleration events emerge as key emission drivers, as shown in Figure . Model predicted CO₂ emissions from older vehicles averaged 14.72 kg/h, approximately 76% higher than those observed in newer vehicles. The trend was likely attributable to progressive engine wear and mechanical efficiency associated with vehicle utilization. The observed variability in emissions among middle-aged vehicles may result from disparities in maintenance, usage intensity, and operating conditions. In port operations, the terrain remained predominantly level with few gradients or minimal altitude variations, thereby minimizing the influence of altitude on emission levels. Emissions were slightly elevated under ascent conditions (10.75 kg/h) relative to level-altitude operation. However, the overall effect of altitude on emission was minimal, fuel demand remained relatively stable over the level terrain of port corridors. The analysis indicated a 39.43% increase in average CO₂ emissions from lower to higher engine displacements, reflecting a substantial rise with engine displacements. Notably, significant emission variability was observed even among vehicles with similar engine size and weight, indicating the influence of operational factors such as idle time, gear usage, vehicle age, and maintenance condition. AFR showed high variability at lower speeds, averaging 21.83 (10–20 km/h) and declining to a stable range at 20–30 km/h Emissions increased progressively across mileage bins, likely due to elevated fuel consumption and reduced efficiency associated with vehicle age and usage. Low-mileage vehicles (15,000–35,000 km) averaged 6.44 kg/h, vehicles with moderate kilometers traveled (35,000–50,000 km) representing 35% increase while heavily used vehicles (>50,000 km), nearly 86% higher emission than low-mileage vehicles.

8.

CO₂ emission predicted trends as influenced under varying real-world operational conditions and vehicle characteristics.

3.4. Feature Dependence SHAP Analysis with Associated Contribution

To assess the influence of individual features on model output, SHAP values were evaluated across the data sets with the variation for specific feature contribution. SHAP assigns an importance value to each feature, reflecting its contribution to the variability in model output, both the degree and direction of each feature impact on model (Figure S10). Significant relationship was observed, with SHAP dependence plot highlighting a varied influence of each operational predictors on emissions, as shown in Figure . Each module shows the distribution of SHAP values for individual features across the data set, indicating potential interactions within the model, with color intensity representing feature statistics. In the distribution, the x-axis represents the feature values, while the y-axis displays the corresponding SHAP values. Such interactions underlines the multifactorial and complex trends under real-world driving conditions.

9.

SHAP-based interpretation of model predicted real-world CO₂ emission highlights the relative importance and directional impact of operational features.

Higher mileage values generally corresponded to slightly positive SHAP values, indicating that vehicle wear over time may contribute to increased emissions. Similarly, SHAP for engine displacement unveiled clustered bands. A high positive and nearly linear trend was observed with both fuel rate and predicted CO₂ emissions (Table S5), indicating that higher fuel usage directly corresponds to elevated emissions. Engine size and age both influenced emissions, with larger engines showing greater SHAP contributions and older vehicles emitting more CO₂, consistent with reduced efficiency and outdated designs. Altitude change contributed minimally, with SHAP centered around zero. However, at extreme AD values (sharp acceleration or deceleration), slight increases in SHAP were observed, suggesting transient operating conditions likely associated with congestion, which may modulate the influence of other features such as mileage, AFR, and engine displacement in port environments.

4. Performance Comparison of Model Outputs

The performance metrics were used to evaluate the predictive accuracy, and generalization of each model, enabling comparison of emission modeling approaches and validate the obtained results. Comparison of RMSE, MAE, and R ² values across models for both train and test data sets is presented in Table . The MLR model demonstrated consistent performance, indicating a reasonable ability to generalize across data sets. However, a minor drop in R ² suggested slight overfitting. LASSO regression improved upon MLR by achieving lower RMSE and MAE values and demonstrated slightly better generalization. Bayesian regression showed better performance than MLR model, yielding lower RMSE values and slightly higher R ² values, indicating that it captured underlying patterns with relatively stable predictive performance. The Random Forest model demonstrated strong performance on the training set, indicating a high capacity to fit the data. However, a notable increase in RMSE and a decline in R ² on the testing set shown a moderate degree of overfitting. Despite this, Random Forest outperformed both MLR, LASSO and Bayesian regression in predictive accuracy for both train and test set, underscoring its suitability for large data applications. The model evaluation, comprising actual–predicted emission (5-fold cross-validation) and residual error distributions, as shown in Figures S6–S9. While the hyperparameter tuned XGBoost model exhibited stable and reliable performance across both data sets. The model significantly enhanced performance over random forest, reducing RMSE by 12.5% and MAE by 7.4%. The XGBoost model showed superior predictive accuracy under real-world conditions, achieving a strong balance between model accuracy and less errors. ,

6. Model Performance Metrics.

data	approach	RMSE	MAE	R 2
MLR
training set	linear regression	3.43	1.61	0.85
testing set		3.87	1.63	0.83
Lasso
training set	regularized linear regression (L1 regularization)	3.41	1.64	0.81
testing set		3.67	1.67	0.79
Bayesian
training set	probabilistic linear regression	2.07	1.79	0.74
testing set		2.14	1.81	0.71
Random Forest
training set	ensemble, bagging, decision tree-based	1.78	0.54	0.94
testing set		1.91	0.57	0.91
XGB Regressor
training set	ensemble, boosting, decision tree-based	1.59	0.51	0.97
testing set		1.63	0.53	0.95

Residual showed minimal bias and homoscedasticity, validating the model reliability in estimating large real-time data sets (Figure ). Achieving both high accuracy and interpretability is critical for freight management and emission characterization. Extending from predicting CO₂ emissions and energy usage, the proposed framework integrates predictive capabilities with spatial analysis, enabling modeling of intensities across locations and facilitating hotspot identification, as shown in Figure . Applicable to traffic regulation in high-impact areas like highways and intersections, this approach enables identification of emission-dense zones that may pose exposure risks to pedestrians. Predicted emissions in the real-world condition closely reflected the observed trends, with model interpretability further strengthening understanding and applicability under complex conditions.

5. Limitation and Future Research

This study evaluates the CO₂ intensity associated with heavy-duty vehicle (HDV) operations within port logistics. Overall, the findings underscores the need to optimize port operations, vehicle performance, and driving behavior to improve efficiency and minimize environmental impacts. Future research may focus on advanced model tuning, integration of zero-emission technologies, and exploration of policy and technological pathways for low-carbon freight systems. The model can be integrated with dispersion modeling frameworks to improve accuracy under extreme conditions, leading to the development of high-precision hybrid models. Further improvements in emission modeling would benefit from detailed spatial data, road networks, and infrastructure information. Moreover, upcoming research should evaluate particulate matter and other harmful gases, given their major role in urban air pollution and health impacts.

6. Conclusions and Policy Implications

In the present study, real-world operational factors such as speed, mileage, altitude, air-fuel ratio, acceleration/deceleration, engine displacement, and weight were found to substantially influence emission outcomes. Vehicle operations resulted in a 64% increase in CO₂ emissions under load, while acceleration phases amplified emissions by 89.5%, underlining the substantial energy requirements of port-based freight operation. Machine learning models were employed to assess emission prediction capabilities. The XGBoost model showed stable and robust predictive performance, with strong R ² metrics indicating that the model effectively captures variance in the data and suitable for real-world use. While Random Forest achieved similar accuracy, it showed signs of overfitting, whereas other regression models offered moderate yet generalizable results. Moreover findings revealed that vehicle usage and age influence emission variability. Vehicles with mileage (50,000–75,000 km) exhibited a 35% increase in emissions, while older fleets (with an average age difference of about 5 years) emitted CO₂ at rates nearly 76% higher (8.31 kg h^–1) than newer vehicles. Additionally, loaded operations showed significantly higher average fuel consumption (5.32 L/h) representing 47% increase compared to off-load conditions. Furthermore, reduced acceleration variability (a < ±2), indicative of smoother driving behavior, contributed to fuel efficiency and emission reductions of up to 23%.

The results underscore the considerable ecological footprint of freight-sector activities. The adoption of regulation from the European Commission marks a paradigm shift in transport decarbonization by introducing CO₂ emission standards for heavy-duty vehicles. The policy underscores the role of data-driven compliance, technological innovation, and regulatory accountability in achieving long-term climate neutrality. This framework serves as a benchmark for developing nations to establish structured emission reduction targets and performance based incentive systems for freight transport. Under the National Clean Air Programme (NCAP), over 108 cities have been identified as nonattainment areas. In response, the government and local authorities have been implementing a range of strategies to curb vehicular pollution in urban regions. Energy-efficient practices, such as the adoption of cleaner fuels and electrification can help reduce the carbon and pollution footprints of logistics operations, as reflected in initiatives like Harit Sagar under the green port guidelines. Given their significant contribution to emissions under real-world driving conditions, it is essential to strengthen vehicle inspection systems to better monitor and track aged, high-mileage, and inadequately maintained heavy-duty vehicles, which often exhibit elevated emission levels. Furthermore, decarbonizing road transportation will require the deployment of advanced solutions, such as solar-powered electrified roads within terminal areas, which can support net-zero emission goals and reduce dependence on fossil fuels.

Data including the code repository for ML in python scripts are available in the model source file at 10.5281/zenodo.17043313

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.5c12136.

• Commodity-wise cargo traffic trends at major Indian ports (Figure S1); specifications of the standard ISO 20-ft container (Table S1); speed-bin-wise fuel consumption and CO₂ emission characteristics under loaded and off-loaded conditions (Table S2); driving phase distribution during transfer operations (Table S3); ANOVA results for speed-bin fuel consumption and CO₂ emissions under load and off-load conditions (Table S4); instantaneous relationships among speed, acceleration/deceleration, fuel consumption, and CO₂ emissions (Figure S2); statistical distribution and normality analysis of fuel consumption and CO₂ emissions (Figure S3); kernel density estimation of fuel consumption (Figure S4); CO₂ emission density distribution with normal fit (Figure.S5); evaluation of MLR, Lasso, Bayesian Ridge, and Random Forest models for actual vs predicted emissions (Figure S6–S9); and SHAP feature importance and contribution analysis (Table S5 and Figure S10) (PDF)

The authors declare no competing financial interest.