HCFF-Net: A Hybrid Contextual Feature Fusion Network for Robust Heavy Equipment Classification in Complex Construction Environments

Abstract

Construction equipment plays a crucial role in performing essential tasks such as construction, demolition, and other maintenance. These machines enable workers to accomplish tasks that would otherwise be extremely difficult or impossible manually, significantly enhancing efficiency and productivity. Heavy construction equipment classification is a critical component of intelligent construction monitoring systems; however, existing vision-based methods often struggle under real-world conditions such as occlusion, background clutter, and scale variation. To address these challenges, this study proposes HCFF-Net, a hybrid contextual feature fusion network designed to enhance classification robustness in complex construction environments. The proposed framework integrates a diverse receptive residual fusion (DRRF) block to capture multi-scale local and global features and a global contextual channel recalibration (GCCR) module to adaptively refine channel-wise representations using contextual information. Unlike conventional feature fusion strategies, HCFF-Net effectively combines structural and contextual features to improve discriminative capability under challenging visual conditions. For performance evaluation, experiments were performed on the publicly available Alberta Construction Image Dataset (ACID). The proposed HCFF-Net achieves a classification accuracy of 90.60% and an F1-score of 90.05% across multiple equipment categories, outperforming state-of-the-art methods, validating its effectiveness for intelligent safety monitoring and management in construction environments.

1. Introduction

The construction industry plays a vital role in the global economy, contributing approximately 13% of the global gross domestic product (GDP) [1]. Heavy construction equipment such as excavators, bulldozers, dump trucks, and loaders forms the backbone of infrastructure development and increasingly supports the integration of artificial intelligence (AI), robotics, and automation into construction operations [2]. These automated technologies aim to improve efficiency and operational monitoring across construction sites [3]. Recent advances in AI and machine learning (ML) have enabled intelligent monitoring systems capable of supporting construction operations throughout the project lifecycle, from planning and execution to maintenance [4]. In particular, deep learning (DL)-based computer vision techniques have demonstrated strong potential for analyzing visual data from construction environments, enabling applications such as activity recognition, safety monitoring, and productivity analysis [5,6].

Reports indicate that heavy equipment plays a central role in construction operations and significantly influences project productivity and workflow management [7]. Effective monitoring of equipment usage is therefore essential for improving operational efficiency and supporting data-driven construction management [8]. With the increasing scale and complexity of modern construction projects, there is a growing need for intelligent monitoring systems capable of automatically analyzing construction site activities using AI-driven solutions [9,10]. In this context, accurate identification of heavy equipment is an important foundational step, as reliable equipment classification enables automated equipment tracking, operational analysis, and digital construction site monitoring.

Accurate identification of construction equipment types is essential because different machines perform distinct operational functions and contribute differently to construction workflows and site management [11]. For example, excavators, dump trucks, and loaders are used for different tasks and therefore require specific monitoring and management strategies. Identifying the type of equipment present on a construction site enables automated equipment tracking, workflow analysis, and data-driven decision-making, which are critical for efficient project management. Automated classification of heavy equipment is therefore a key component of intelligent construction monitoring systems, as it provides reliable input for downstream analysis tasks such as equipment utilization assessment and progress tracking in complex environments [6]. Beyond the technical contribution, this capability has significant practical importance, as effective monitoring of equipment usage is directly associated with improved project productivity and operational efficiency [6,7]. As construction sites become increasingly complex, automated visual analysis using deep learning reduces reliance on manual supervision and supports timely decision-making under dynamic conditions [12]. Furthermore, such capabilities are essential for modern smart construction applications, including AI-driven monitoring systems and digital twin-based platforms, where accurate equipment identification contributes to reliable and real-time site intelligence [8,10,13]. Therefore, robust heavy equipment classification is a fundamental requirement for developing intelligent, safe, and automated construction management systems.

Although recent studies have demonstrated promising progress, several challenges remain. Existing approaches often experience performance degradation when applied to real construction environments due to occlusion, background clutter, illumination variations, and scale changes [3,4]. Furthermore, many existing methods are evaluated on controlled datasets, limiting their applicability to dynamic real-world construction environments. In addition, prior studies have largely focused on object detection tasks, such as personal protective equipment recognition [14], unsafe behavior detection [15], and load monitoring [16,17], rather than detailed classification of heavy equipment categories. These limitations highlight the need for robust classification frameworks specifically designed for complex construction environments.

This study aims to develop a robust deep learning framework for accurate heavy equipment classification in complex construction environments. The primary research objective is to design a model capable of effectively handling real-world challenges such as occlusion, background clutter, and scale variation. Based on this objective, the key research question is as follows: How can deep learning architectures be designed to improve classification performance under complex construction site conditions? To address this question, the study hypothesizes that integrating multi-scale receptive feature extraction with contextual feature recalibration can significantly enhance classification accuracy and robustness under challenging visual conditions. As a case study, the proposed framework is evaluated on the publicly available Alberta Construction Image Dataset (ACID) [18], which contains diverse construction equipment images captured under realistic site conditions. The scope of this work is limited to equipment classification and does not include worker behavior analysis or activity recognition. While safety and operational applications are discussed, they are considered as potential downstream applications rather than directly evaluated outcomes.

To systematically identify relevant studies, a structured literature review was conducted focusing on research related to construction heavy equipment analysis using machine learning and deep learning techniques. The literature search was performed using major scientific databases including Scopus, Web of Science, Google Scholar, and ScienceDirect. The search process used combinations of keywords such as heavy equipment classification, construction equipment recognition, construction vehicle detection, deep learning in construction, construction computer vision, equipment activity recognition, and vision-based construction monitoring. The initial search results were screened based on the relevance of titles and abstracts. Studies were included if they applied ML or DL methods to construction equipment analysis, used visual data from construction environments, addressed equipment classification, detection, or activity recognition problems, and provided experimental validation on construction-related datasets. The selected studies were then categorized into machine learning approaches and deep learning approaches to better illustrate the evolution of intelligent construction monitoring methods. The review indicates that while detection and activity recognition have been widely studied, relatively limited research focuses specifically on heavy equipment classification, which motivates the focus of this study. A comparison of the selected studies, including their strengths and limitations, is presented in Table 1.

1.1. Machine Learning Approaches

Several studies have applied machine learning (ML) techniques to address construction-related problems such as equipment management, productivity analysis, safety assessment, and cost estimation. Shehadeh et al. [19] proposed three ML algorithms to estimate the residual value of heavy construction equipment. Among these, the modified decision tree model achieved the best performance for supporting decisions related to equipment purchasing, replacement, and disposal. Their study utilized six datasets containing eleven variables and more than 60,000 equipment records. Choi et al. [20] applied a random forest model to identify construction workers at high risk of fatal accidents using national safety data, demonstrating the potential of ML in improving construction safety management. Similarly, Chen et al. [21] developed a Bayesian optimization-based gradient boosting model to predict tunnel deformation. Their model incorporated SHAP analysis for interpretability and demonstrated strong predictive accuracy using data from the Wuhan Metro project. Eriksson and Ghabcheloo [22] evaluated five ML models across multiple datasets and found that multilayer perceptron and CNN-based models achieved superior performance compared to other approaches.

Sadatnya et al. [23] proposed an ML-based framework for estimating crew productivity across different construction projects by incorporating operational progress, weather conditions, resource availability, and crew composition. Their framework demonstrated practical benefits for project planning and scheduling. Boyko and Lukash [24] developed an ML-based methodology for predicting heavy equipment costs based on equipment specifications and geographic attributes, demonstrating the usefulness of ML in construction cost forecasting. Momade et al. [25] modeled construction labor productivity using support vector machines and random forest models, achieving prediction accuracies above 90%. Gurmu and Ongkowijoyo [26] developed a logistic regression model to evaluate construction labor productivity based on human resource management practices, showing that improved HR practices can reduce delays and improve productivity. To address data integration challenges, You and Wu [27] proposed an enterprise-integrated data platform to improve interoperability and decision support by combining diverse datasets, demonstrating improvements in project management efficiency. Kim et al. [28] use interpretable machine learning to estimate damage in construction crane accidents. Using a dataset of 701 accident cases, the model identifies key influencing factors, such as boom length and outrigger deployment, while highlighting the importance of text-based features. Song et al. [29] proposes a regression-based model to evaluate accident risk in construction equipment by analyzing utilization rates, subcontractor types, and construction costs.

Despite these contributions, most ML-based approaches rely on structured tabular data rather than visual data and therefore have limited applicability to automated visual monitoring systems in construction environments. This limitation highlights the growing need for vision-based approaches capable of analyzing complex construction site imagery.

1.2. Deep Learning Approaches

Deep learning (DL) has gained significant attention in construction research due to its ability to automatically learn hierarchical feature representations from visual data. Mannem et al. [30] introduced the WATLAS framework, which applies adaptive transfer learning using InceptionV3 to classify 15 construction object categories using only 5% labeled data, achieving approximately 90% classification accuracy. Shin et al. [31] applied a 3D residual neural network for excavator activity recognition, categorizing operations into working and non-working activities and achieving an F1-score of 81.8%. DL methods have also been widely applied to safety monitoring. Sun et al. [32] proposed a housekeeping change detection network (HCDN) to improve construction site safety, achieving 89.32% accuracy and an F1-score of 86.67%. Similarly, other studies have applied DL methods for automated waste classification to improve recycling efficiency in construction environments [33]. In productivity monitoring applications, Liu et al. [34] proposed a vision-based framework for classifying full and empty dump trucks, while Chen et al. [35] developed a CNN-based method for recognizing excavator activities and estimating productivity, achieving 87.6% activity recognition accuracy and 83% productivity estimation accuracy.

DL techniques have also been applied to safety analysis and monitoring tasks. Uhm et al. [36] utilized large multimodal models to analyze construction accident videos, significantly reducing manual analysis effort. Yang et al. [37] proposed a transformer-based model to detect unsafe worker behaviors from video data, achieving an average precision of 88.7%. For site monitoring applications, Kim and Chi [38] developed a multi-camera framework for tracking earthmoving productivity, achieving 97.6% identification accuracy. Arabi et al. [39] proposed a DL-based approach for real-time construction vehicle detection in dynamic site conditions. Shi et al. [40] combined semi-supervised learning and data augmentation to achieve 81.1% mAP on the ACID while using only 3% labeled data. Nath et al. [41] applied transfer learning with VGG-16 for single- and multi-label classification of construction objects, achieving approximately 90% and 85% accuracy. Yamany et al. [42] proposed a convolutional neural network (CNN)-based approach for the classification of heavy construction equipment and demonstrating the effectiveness of deep learning models for automated equipment recognition in construction environments, respectively. Shin et al. [43] propose a benchmark model for the automated detection and classification of a wide range of heavy construction equipment, addressing limitations of existing datasets and models. Eum et al. [44] proposed a heavy equipment detection method that integrates a transformer-based backbone into the YOLO framework, improving feature representation and enhancing detection performance in complex construction site conditions. Helian et al. [45] propose a computer vision-based method that uses depth maps and a customized Faster R-CNN model to estimate excavator bucket fill, achieving high accuracy and enabling real-time productivity monitoring in construction operations. Kim et al. [46] proposed a vision-based excavator activity classification method using a CNN-BiLSTM framework to classify excavator activities by analyzing visual information from consecutive video frames. Elelu et al. [47] proposed an audio-based framework using a convolutional recurrent neural network (CRNN) to classify and localize construction equipment sounds for collision hazard prevention by analyzing multi-channel audio signals. Kim et al. [48] propose an SRGAN-enhanced cascade learning framework for detecting and classifying unsafe operations of heavy construction machinery, which improves low-quality CCTV footage resolution and enables more accurate action recognition in complex construction environments.

Table 1 summarizes representative studies in construction equipment analysis, highlighting their methodologies, strengths, and limitations. This comparison provides important context for identifying remaining research challenges and demonstrates the research gap discussed in the next section.

1.3. Research Gap

Based on the reviewed literature (from Section 1.1 and Section 1.2), there remains a lack of robust DL frameworks specifically designed for heavy construction equipment classification in complex real-world environments. Although existing studies demonstrate the effectiveness of DL in construction monitoring, most research focuses on detection and activity recognition. Furthermore, many studies evaluate limited datasets and do not fully address challenges such as occlusion, background clutter, and environmental variability commonly found in real construction environments. Limited attention has also been given to improving classification robustness under realistic construction conditions. Therefore, there is a need for a DL-based classification framework capable of addressing these challenges while improving classification performance across diverse equipment categories. To address this identified research gap, this study proposes a DL-based heavy equipment classification framework designed for complex construction environments. The proposed method leverages enhanced feature extraction and contextual feature fusion strategies to improve classification robustness for heavy equipment categories under challenging visual conditions.

The remainder of this paper is organized as follows. Section 2 presents the proposed HCFF-Net architecture and experimental setup. Section 3 presents the experimental results. Section 4 discusses the findings and limitations. Finally, Section 5 concludes the paper.

Table 1. Comparison of previous studies on construction heavy equipment.

Category	Method	Dataset	Strength	Limitation
Machine learning	Light GBM, XGBoost [19]	Heavy equipment datasets	Predict residual value of heavy equipment	-Depends on data quality and quantity -Performance improves with better data collection
	Random Forest [20]	National accident database	Identifies high-risk workers	Limited by historical data and missing variables due to privacy constraints
	BO-NGBoost [21]	Wuhan Metro data	Forecasts tunnel deformation	Limited sample size and geographical scope
	Hyperparameter-tuned ML models [23]	Daily work reports	Estimate crew productivity effectively	Low data quality and high computational cost
	ML Regression models [24]	Self-collected equipment data	Predict equipment cost trends	-Sensitive to market -Relies on inconsistent data
	Logistic Regression [26]	39 contractors work on multistory building projects	Predicts construction labor productivity through HRM practices	-Findings limited to Australian projects -Excludes union data
	ML + SHAP [28]	Crane accident dataset	Informed safety decision-making	Limited to structured data
	Regression model [29]	Equipment cost data	Quantitative risk prediction	-Ignores equipment subcategories -Unvalidated normal cost assumptions
Deep learning	WATLAS [30]	Construction site images	Performs well on imbalanced data with limited labels	-Single-region dataset limits generalization -Needs expert labeling and tuning
	3D ResNet [31]	Excavator videos	Activity recognition via transfer learning and SHAP-based camera analysis	-Limited to excavators -May not generalize to other equipment
	HCDN [32]	Housekeeping-CCD dataset (construction site images)	Detects housekeeping changes accurately in real site images	-Device and site-specific -Slower training and limited scalability
	CNN & Swin Transformer [33]	RGB images of end-of-life plastic waste	High accuracy with low-cost RGB images	-Limited to four types of plastic waste -RGB images are not enough for detailed classification
	CNN & Transfer learning [34]	Field images of trucks	Compares multiple architectures Transfer learning improves accuracy	-Only works for uncovered trucks in open sites -Limited generalizability
	CNN-based framework [35]	Project surveillance videos	Automates equipment tracking and reduces manual monitoring	-Limited to specific sites and excavator types -Performance affected by detection and lighting
	LMM + Graph RAG [36]	Accident video footage	Automates safety analysis using multimodal reasoning	-Designed for specific accident types -Limited contextual transferability
	STR-Transformer [37]	Unsafe action video dataset	Monitors worker safety and action effectively	-Requires large dataset -Limited to predefined action classes
	Multi-camera vision DL [38]	Earthmoving project video	Enables productivity monitoring via multiple cameras	-Relies on fixed multi-camera setup -Poor adaptability to other layouts
	SSD-MobileNet [39]	Dynamic jobsite footages	High real-time accuracy	-Small dataset -Tested on limited vehicle types and conditions
	Semi-supervised DL [40]	ACID	Enhances site monitoring using data augmentation	-Requires careful tuning -Performance drops with noisy images
	VGG-16 [41]	Construction site images	Handles single and multi-label classification	-Small dataset -Transfer learning misses domain features
	CNN model [42]	Heavy equipment images	Automated heavy equipment classification	-No advanced feature fusion for complex environments
	CNN model [43]	Custom heavy equipment dataset	Multi-class detection	Limited generalization to unseen classes
	YOLOv10 + Transformer [44]	Construction site images	Heavy equipment detection	-Needs more diverse training data -Requires better domain adaptation
	Faster R-CNN [45]	Excavator-specific dataset	Bucket filling estimation	Task-specific and low generalization
	CNN–BiLSTM [46]	Excavator activity video dataset	Classify excavator activities	Limited to excavator-specific activities with moderate generalization capability
	CRNN [47]	Equipment sound dataset	Equipment sound-based classification	Limited realism and noise representation in synthetic data
	SRGAN Network [48]	Custom heavy equipment dataset	Enhancing safety in heavy equipment operations	Limited equipment types and dataset
	Proposed HCFF-Net	ACID	-Good generalization across equipment types -DRRF + GCCR improves feature quality	-Requires GPU for training -Needs diverse data for rare equipment

Table 1. Comparison of previous studies on construction heavy equipment.

Category	Method	Dataset	Strength	Limitation
Machine learning	Light GBM, XGBoost [19]	Heavy equipment datasets	Predict residual value of heavy equipment	-Depends on data quality and quantity -Performance improves with better data collection
	Random Forest [20]	National accident database	Identifies high-risk workers	Limited by historical data and missing variables due to privacy constraints
	BO-NGBoost [21]	Wuhan Metro data	Forecasts tunnel deformation	Limited sample size and geographical scope
	Hyperparameter-tuned ML models [23]	Daily work reports	Estimate crew productivity effectively	Low data quality and high computational cost
	ML Regression models [24]	Self-collected equipment data	Predict equipment cost trends	-Sensitive to market -Relies on inconsistent data
	Logistic Regression [26]	39 contractors work on multistory building projects	Predicts construction labor productivity through HRM practices	-Findings limited to Australian projects -Excludes union data
	ML + SHAP [28]	Crane accident dataset	Informed safety decision-making	Limited to structured data
	Regression model [29]	Equipment cost data	Quantitative risk prediction	-Ignores equipment subcategories -Unvalidated normal cost assumptions
Deep learning	WATLAS [30]	Construction site images	Performs well on imbalanced data with limited labels	-Single-region dataset limits generalization -Needs expert labeling and tuning
	3D ResNet [31]	Excavator videos	Activity recognition via transfer learning and SHAP-based camera analysis	-Limited to excavators -May not generalize to other equipment
	HCDN [32]	Housekeeping-CCD dataset (construction site images)	Detects housekeeping changes accurately in real site images	-Device and site-specific -Slower training and limited scalability
	CNN & Swin Transformer [33]	RGB images of end-of-life plastic waste	High accuracy with low-cost RGB images	-Limited to four types of plastic waste -RGB images are not enough for detailed classification
	CNN & Transfer learning [34]	Field images of trucks	Compares multiple architectures Transfer learning improves accuracy	-Only works for uncovered trucks in open sites -Limited generalizability
	CNN-based framework [35]	Project surveillance videos	Automates equipment tracking and reduces manual monitoring	-Limited to specific sites and excavator types -Performance affected by detection and lighting
	LMM + Graph RAG [36]	Accident video footage	Automates safety analysis using multimodal reasoning	-Designed for specific accident types -Limited contextual transferability
	STR-Transformer [37]	Unsafe action video dataset	Monitors worker safety and action effectively	-Requires large dataset -Limited to predefined action classes
	Multi-camera vision DL [38]	Earthmoving project video	Enables productivity monitoring via multiple cameras	-Relies on fixed multi-camera setup -Poor adaptability to other layouts
	SSD-MobileNet [39]	Dynamic jobsite footages	High real-time accuracy	-Small dataset -Tested on limited vehicle types and conditions
	Semi-supervised DL [40]	ACID	Enhances site monitoring using data augmentation	-Requires careful tuning -Performance drops with noisy images
	VGG-16 [41]	Construction site images	Handles single and multi-label classification	-Small dataset -Transfer learning misses domain features
	CNN model [42]	Heavy equipment images	Automated heavy equipment classification	-No advanced feature fusion for complex environments
	CNN model [43]	Custom heavy equipment dataset	Multi-class detection	Limited generalization to unseen classes
	YOLOv10 + Transformer [44]	Construction site images	Heavy equipment detection	-Needs more diverse training data -Requires better domain adaptation
	Faster R-CNN [45]	Excavator-specific dataset	Bucket filling estimation	Task-specific and low generalization
	CNN–BiLSTM [46]	Excavator activity video dataset	Classify excavator activities	Limited to excavator-specific activities with moderate generalization capability
	CRNN [47]	Equipment sound dataset	Equipment sound-based classification	Limited realism and noise representation in synthetic data
	SRGAN Network [48]	Custom heavy equipment dataset	Enhancing safety in heavy equipment operations	Limited equipment types and dataset
	Proposed HCFF-Net	ACID	-Good generalization across equipment types -DRRF + GCCR improves feature quality	-Requires GPU for training -Needs diverse data for rare equipment

2. Experimental Setup and Methods

For this study, the ACID [18] was selected because it is specifically designed for construction equipment analysis and contains real-world construction site images with detailed annotations. Unlike general-purpose datasets such as ImageNet [49], which contain construction equipment as part of broader object categories, ACID focuses specifically on construction machinery, making it more suitable for evaluating equipment classification methods. Furthermore, the dataset contains significant variations in viewpoint, illumination, occlusion, and background complexity, which are critical challenges in real construction environments. These characteristics make ACID particularly appropriate for evaluating the robustness and generalization capability of deep learning models under realistic operating conditions.

ACID contains ten predefined heavy equipment categories (backhoe loader, compactor, cement truck, dozer, dump truck, excavator, grader, mobile crane, tower crane, and wheel loader) comprising a total of approximately 10,000 images, as given in

Table 2. Heavy equipment classes with sample numbers in the ACID.

No. of Classes	Heavy Equipment Type	Samples
1	Backhoe Loader	1300
2	Compactor	979
3	Cement Truck	730
4	Dozer	959
5	Dump Truck	729
6	Excavator	1506
7	Grader	1240
8	Mobile Crane	938
9	Tower Crane	295
10	Wheel Loader	1324
-	Total	10,000

. In this study, all available classes were utilized to ensure a comprehensive evaluation of the proposed model rather than selecting a subset of categories. The use of all dataset classes avoids selection bias and allows fair comparison with other methods. Furthermore, these equipment types represent commonly used machinery in construction projects and provide sufficient inter-class diversity for evaluating classification robustness. The dataset was split into 70% for training, 10% for validation, and 20% for testing. This split was adopted to ensure a balanced trade-off between model training and unbiased performance evaluation. Representative samples for each class are shown in Figure 1 and the number of samples per class is summarized in Table 2.

The hardware setup consisted of an Intel^® Core™ i7-7700 processor operating at 2.80 GHz, 16 GB of random-access memory (RAM), and an NVIDIA GeForce GTX 1050 graphics processing unit (GPU) with 4 GB memory. All experiments were implemented in Python using the PyTorch framework [50], which was selected due to its dynamic computation graph and flexibility for research-oriented deep learning development. Other frameworks, such as TensorFlow, also provide strong capabilities for large-scale deployment [51]. However, PyTorch was chosen because it better supports rapid architectural experimentation and customization required in this study. PyCharm (version 2025.1.1.1) was used as the development environment due to its advanced debugging capabilities, efficient project management features, and seamless integration with Python (version 3.12.7), which facilitated efficient experimentation and improved implementation reproducibility. It should be noted that the choice of development environment does not affect the model design but supports structured experimentation and implementation reliability.

2.1. Proposed Method and Development Workflow

The development of the proposed system is described in a sequential step-by-step manner, as illustrated in Figure 2. In step 1, the ACID is prepared and organized for the heavy equipment classification task. The dataset is divided into training, validation, and testing subsets, where the training set is used to learn model parameters, the validation set is used to monitor the learning process and optimize model settings, and the testing set is reserved for final evaluation on unseen images. In step 2, each construction site image is resized to a fixed input size and provided to the proposed HCFF-Net model. In step 3, The processed input image is then simultaneously passed into two parallel branches: the Dense Model and the diverse receptive context-aware network (DRC-Net). The Dense Model is responsible for extracting fine-grained visual features, while the DRC-Net extracts hierarchical and contextual features from the same input image. In step 4, the features obtained from the Dense Model and the DRC-Net are combined through a feature fusion process. This step integrates both local structural and global contextual representations. The fused features are then passed through a global average pooling layer followed by a classification layer to predict the final heavy equipment category. Finally, in step 5, unseen images are provided during the testing phase, where the trained HCFF-Net generates the final classification results and predicts the equipment category.

The overall training and testing workflow of the proposed method is shown in Figure 3. The proposed HCFF-Net works in two main phases: training and testing. During the training phase, HCFF-Net is trained using labeled data, while in the testing phase, the trained model is evaluated on an unseen portion of the dataset. The HCFF-Net consists of two parallel convolutional neural network branches: DRC-Net and a Dense Model. During training, the input images are processed by a Dense Model to extract rich and discriminative features representations. In parallel, the same input images are fed into the DRC-Net, which is composed of four sequential stages. Each stage consists of a convolutional block followed by a proposed DRRF block. The DRRF block enhances multi-scale feature representation by aggregating residual features extracted through dilated convolutions with different receptive fields. At the final stage of the DRC-Net, the proposed GCCR block is applied to adoptively reweight feature responses by modeling global contextual information through channel-wise attention. The resulting features from both branches are then forwarded to the classification stage for model optimization. During the testing phase, unseen images from the dataset are directly fed into the trained HCFF-Net. The network utilizes the learned contextual and hierarchical features to generate the final classification results. This two-stage workflow helps the model learn features effectively during training and perform reliably during testing.

2.2. HCFF-Net Structure

The detailed structure of the proposed HCFF-Net is shown in Figure 4. While Figure 2 presents the overall development workflow of the system, Figure 4 explains how the internal architecture of HCFF-Net is organized to perform heavy equipment classification. The main idea of the proposed architecture is to combine two complementary branches so that both fine-grained local features and high-level contextual features can be learned effectively. This design is particularly important for construction site images, where equipment categories often appear under complex conditions such as occlusion, background clutter, and large-scale variation.

Initially, the proposed network HCFF-Net obtains the input image $\left(I\right) \in {\mathbb{R}}^{224 \times 224 \times 3}$ pixel which is simultaneously fed into two networks: Dense Model and DRC-Net. The backbone of the DRC-Net is based on the ConvNeXt-Small architecture [52] while the Dense Model is based on DenseNet-121 [53]. To justify the selection of backbone networks, various pretrained models were evaluated on the dataset, and the results are reported in Section 3. Among these models, ConvNeXt-S achieved the highest classification accuracy, which motivated its selection as the primary backbone. To further improve feature learning, the architecture was extended by integrating the proposed DRRF and GCCR modules and designed DRC-Net. The effectiveness of these architectural improvements is validated through the ablation study results shown in Table 4. DenseNet121 was selected as a complementary branch because it achieved the second-highest accuracy among the evaluated models, as shown in Table 5. Due to its dense feature reuse capability, it helps preserve fine-grained information. By combining DRC-Net with the Dense Model, the proposed HCFF-Net benefits from complementary feature representations. Therefore, a hybrid architecture was designed to concatenate features from both branches, aiming to improve overall classification performance, as demonstrated in

Table 3. Performance comparisons of each sub-network and proposed HCFF-Net.

Case	Accuracy	Precision	Recall	F1-Score
Dense Model	88.41	88.18	87.16	87.67
DRC-Net	90.10	89.78	89.29	89.53
HCFF-Net	90.60	90.60	89.52	90.05

.

In the DRC-Net, the input image $\left(I\right) \in {\mathbb{R}}^{224 \times 224 \times 3}$ is first passed through an initial convolution layer, producing features maps $F\in {\mathbb{R}}^{56 \times 56 \times 96}$. These features are processed by the first stage of the DRC-Net, which focuses on extracting low-level spatial features. This stage consists of three convolutional layers followed by the first DRRF block. The DRRF block employs dilated convolutions with multiple receptive fields to enhance contextual awareness and emphasize informative regions, while preserving the feature map size $F\in {\mathbb{R}}^{56 \times 56 \times 96}$. The features are then forwarded to the second stage, where a stride-2 convolutional layer performs spatial down sampling and increases the channel dimension to 192, resulting in feature maps of size $F\in {\mathbb{R}}^{28 \times 28 \times 192}$. This stage also comprises three convolutional layers, followed by the second DRRF block to further refine mid-level feature representations of heavy equipment. Next, the refined features are passed to the third stage of DRC-Net, which applies additional spatial down sampling and a deeper stack of convolutional layers followed by the third DRRF block. This stage produces high-level semantic feature maps of size $F\in {\mathbb{R}}^{14 \times 14 \times 384}$. Subsequently, the fourth stage further processes the features using three convolutional layers followed by a fourth DRRF block. The resulting feature maps are then passed through the proposed GCCR block, which adaptively recalibrates feature responses using channel-wise attention. The final output of the DRC-Net has a feature size of $F\in {\mathbb{R}}^{7 \times 7 \times 768}$.

In parallel, the Dense Model processes the same input images of size $224 \times 224 \times 3$. The extracted features are sequentially passed through dense block 1 to dense block 4, producing an output feature map of size $F\in {\mathbb{R}}^{7 \times 7 \times 1024}.$ Finally, the output feature maps from DRC-Net $F\in {\mathbb{R}}^{7 \times 7 \times 768}$ and Dense Model $F\in {\mathbb{R}}^{7 \times 7 \times 1024}$ are concatenated and pass through a global average pooling layer, resulting in a fused feature vector of size $F\in {\mathbb{R}}^{1 \times 1 \times 1792}$. This representation is fed into a linear classifier layer to predict the final classification results across 10 heavy equipment classes. The study provides a detailed explanation of the proposed novel blocks in the following subsections.

2.3. DRRF

Existing multi-scale fusion approaches, such as Atrous Spatial Pyramid Pooling (ASPP) [54], capture contextual information through multiple parallel branches with predefined dilation rates (e.g., 6, 12, and 18). The resulting feature maps are commonly aggregated via concatenation followed by dimensionality reduction. However, the use of fixed and sparsely distributed dilation rates may limit the ability to capture fine-grained local variations, which are essential for distinguishing visually similar object categories. In contrast, the proposed DRRF module improves multi-scale representation by integrating diverse receptive fields through residual fusion rather than simple parallel concatenation. Specifically, DRRF employs convolutional blocks combined with normalization and GELU activation to progressively extract hierarchical features. These features are then fused through residual connections, allowing both local detail information and broader contextual information to be preserved simultaneously. Furthermore, unlike ASPP, which is typically applied at a single high-level feature stage, the proposed DRRF block is integrated at multiple stages in the proposed model to enhance feature learning at different semantic levels. This multi-stage integration allows the network to capture both low-level texture details and high-level semantic structures, improving robustness under challenging conditions such as occlusion, scale variation, and background clutter.

As illustrated in Figure 5, the DRRF block consists of two parallel convolutional branches. The first branch employs a $3\times 3$ dilated convolution with a dilation rate of 5, while the second branch utilizes a standard $3\times 3$ convolution. In both branches, the convolution operation is followed by a batch normalization (BN) layer and a Gaussian Error Linear Unit (GELU) activation function. This dual-branch design enables the DRRF block to jointly capture fine-grained local features through standard convolution and enriched contextual information through dilated convolution, without compromising spatial resolution. Let the input feature tensor to the DRRF block be $F^{DRRF\left(I\right)}$ $\in$ ${\mathbb{R}}^{C \times H \times W}$. The output feature tensor $F^{DRRF\left(O\right)}\in {\mathbb{R}}^{C\times H\times W}$ is obtained through a residual fusion mechanism, as defined in Equation (1):

$$F^{DRRF\left(O\right)}= F^{DRRF\left(I\right)}+ F^{DRRF\left(C\right)}$$

(1)

Here, $F^{DRRF\left(C\right)}$ represents the fused feature map generated by concatenating the outputs of the two convolutional branches along the channel dimension, as expressed in Equation (2):

$$F^{DRRF\left(C\right)} = F^{DRRF\left(A\right)} \copyright F^{DRRF\left(B\right)}$$

(2)

The feature map $F^{DRRF\left(A\right)}$ produced by the dilated convolution branch is computed as follows:

$$F^{DRRF\left(A\right)} =GELU\left(BN\left(Dil Covn\left(K_{3 \times 3}, F^{DRRF\left(I\right)}\right)\right)\right)$$

(3)

The feature map $F^{DRRF\left(C\right)}$ obtained from the standard convolution branch is given by

$$F^{DRRF\left(B\right)}=GELU \left(BN\left(Conv\left(K_{3\times 3}, F^{RFIE\left(I\right)}\right)\right)\right)$$

(4)

Through this dual-branch residual fusion strategy, the DRRF block effectively integrates local structural details and contextual information, thereby enhancing feature representation capability for subsequent stages of the network.

2.4. GCCR

Existing attention mechanisms, such as the Squeeze-and-Excitation (SE) block [55], model channel importance by applying global average pooling followed by fully connected layers to capture channel-wise dependencies. Although effective in modeling global semantic relationships, this approach primarily relies on global feature statistics and does not explicitly enhance contextual feature interactions, which are essential for distinguishing visually similar categories. Additionally, SE typically employs a fixed ReLU activation, whereas the proposed GCCR module adopts PReLU activation, enabling adaptive learning of activation parameters and greater flexibility in feature representation. Similarly, the Convolutional Block Attention Module (CBAM) [56] integrates both channel and spatial attention through global pooling operations followed by sequential attention refinement. While CBAM improves feature weighting, its attention mechanism is largely based on global aggregation and sequential processing. As a result, it may not sufficiently capture subtle contextual variations required for fine-grained classification tasks, such as heavy equipment recognition, where occlusion, background clutter, and scale variations are common.

In contrast, the proposed Global Contextual Channel Recalibration (GCCR) module is specifically designed to enhance contextual channel interactions through a more efficient and adaptive architecture. Instead of relying on fully connected layers as in SE, GCCR utilizes global contextual aggregation followed by a lightweight 1 × 1 convolution and batch normalization. This design reduces parameter complexity while improving feature stability. The recalibrated features are further refined using learnable channel scaling, adaptive PReLU activation, and dropout regularization, enabling the network to emphasize informative channels while improving generalization and reducing overfitting. Unlike CBAM, which applies channel and spatial attention sequentially, GCCR focuses on contextual channel refinement with integrated regularization, enhancing discriminative feature learning without increasing architectural complexity. Moreover, GCCR incorporates adaptive activation (PReLU) and dropout regularization, which are not present in SE or CBAM, thereby improving robustness.

GCCR is applied after the final stage of the DRC-Net to recalibrate feature responses without altering spatial resolution as shown in Figure 6. Let the input feature tensor to the GCCR block be $F^{GCCR\left(I\right)} \in {\mathbb{R}}^{C \times H \times W}$. First, the global contextual information is aggregated using average pooling, producing an intermediate feature map $P,$ as defined in Equation (5):

$$P=Pool \left(K_{3 \times 3, } F^{GCCR\left(I\right)}\right)$$

(5)

The pooled features are then passed through a $1\times 1$ convolution, followed by batch normalization, a sigmoid activation, and parametric rectified linear unit (PReLU) activation to generate channel-wise attention weights, as expressed in Equation (6):

$$A= Sigmoid\left(BN\left(Conv\left(K_{1 \times 1, } P\right)\right)\right)\times F^{GCCR\left(I\right)}$$

(6)

To improve generalization, a dropout layer with a rate of $0.5$ is applied to the recalibrated features. Finally, the output feature tensor of the GCCR block is obtained through residual learning, as defined in Equation (7):

$$F^{GCCR\left(o\right)} =Dropout\left(0.5, PReLU\left(A\right)\right)$$

(7)

2.5. Statistical Analysis Using McNemar’s Test

McNemar’s test [57] is a non-parametric statistical method used to compare the performance of two classification models evaluated on the same dataset. In this study, McNemar’s test is employed to assess whether there is a statistically significant difference in classification performance between the proposed HCFF-Net and the baseline model (ConvNeXt-S). The test is based on the number of samples that are misclassified differently by the two models. Let b denote the number of samples correctly classified by HCFF-Net but misclassified by the baseline model and c denote the number of samples correctly classified by the baseline model but misclassified by HCFF-Net. The McNemar test statistic is computed as

$$x^2={\displaystyle \frac{{\left(\left( b-c \right)-1\right)}^2}{b+c}}$$

(8)

3. Experimental Results

The proposed HCFF-Net was trained using a learning rate of 0.0001 and a batch size of 8. The model was trained for seven epochs, and an early stopping strategy was applied to prevent overfitting, where validation accuracy was used as the stopping criterion. The Adam optimizer [58] and cross-entropy loss function [59] were used during training. The total training time was approximately one hour on a hardware system described in Section 2. First, the Dense Model was trained independently, and its weight was saved. Due to dense feature reuse and strong gradient flow, the model learns the training data very efficiently due to it continuing to improve on the training data after the early epochs. Its training curves are shown in Figure 7a. Next, the DRC-Net incorporating the proposed convolutional blocks at each stage was trained, and its weights were saved. The DRC-Net introduces DRRF and GCCR blocks along with attention mechanisms, which increase model complexity. This results in slower convergence but encourages the learning of more discriminative and structured features. Its training curves are shown in Figure 7b. Finally, the proposed HCFF-Net was trained. The ensemble HCFF-Net exhibits rapid convergence with near-zero training loss and perfect training accuracy due to the use of pretrained constituent networks. Validation accuracy remains consistently high, while validation loss stabilizes after minor early fluctuations, indicating improved generalization and reduced variance. This behavior shows that the HCFF-Net effectively combines features from the Dense Model and DRC-Net without introducing additional overfitting. The corresponding training curves are shown in Figure 7c.

3.1. Evaluation Merics

Four evaluation metrics, $Accuracy$, $Precision$, $Recall$, and $F1$-score, were used to assess the performance of the proposed model. These metrics were computed based on the values of true positives $\left(TP\right)$, true negatives $\left(TN\right)$, false positives $\left(FP\right)$, and false negatives $\left(FN\right)$. $TP$ and $TN$ represent correct predictions for positive and negative samples, respectively, whereas $FP$ denotes instances where the model incorrectly predicts a positive class for a negative sample, and $FN$ refers to cases where a positive sample is incorrectly classified as negative, and n denotes the total number of classes in the dataset.

The mathematical formulations of these evaluation metrics are presented in Equations (9)–(12):

$$Accuracy= {\displaystyle \frac{1}{n}} {Ʃ}_{i=1}^n {\displaystyle \frac{{TP}_i+ {TN}_i}{{TP}_i+ {TN}_i+ {FP}_i+ {FN}_i}}$$

(9)

$$Precision={\displaystyle \frac{1}{n}} {Ʃ}_{i=1}^n {\displaystyle \frac{{TP}_{i }}{{TP}_{i }+{FP}_i }}$$

(10)

$$Recall={\displaystyle \frac{1}{n}} {Ʃ}_{i=1}^n {\displaystyle \frac{{TP}_{i }}{{TP}_{i }+{FN}_i }}$$

(11)

$$F1\text{-}score=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(12)

3.2. Ablation Study

In the ablation study, the efficiency of the proposed model was evaluated to better understand the role of proposed HCFF-Net. The results of these cases are presented in Table 3. Each case represents a different network setup, starting from a basic configuration and gradually adding key components. The performance of the baseline Dense Model, the standalone DRC-Net, and the final HCFF-Net integrating both components was progressively analyzed. The baseline Dense Model achieved an accuracy of 88.41%, serving as a reference for subsequent comparisons. When the DRC-Net was applied independently, noticeable improvements were observed across all evaluation metrics, with accuracy increasing to 90.10% and F1-score to 89.53%. This demonstrates the capability of the DRC block to enhance feature representation by effectively capturing discriminative patterns.

The proposed HCFF-Net, which combines the Dense Model with the DRC-Net, achieved the best overall performance. Specifically, HCFF-Net attained an accuracy of 90.60%, precision of 90.60%, recall of 89.52%, and an F1-score of 90.05%. Compared to the baseline Dense Model, HCFF-Net improved accuracy by 2.19% and F1-score by 2.38%, confirming the complementary benefits of feature fusion between the two blocks. These results clearly indicate that each component contributes positively to the final performance, and their integration leads to more robust and discriminative feature learning. The ablation study validates the effectiveness of the proposed HCFF-Net design for heavy construction equipment classification.

Another ablation study was conducted to analyze the contribution of each proposed block within the DRC-Net architecture. As presented in

Table 4. Performance comparisons with or without the proposed blocks.

Case	DRRF	DRRF	DRRF	GCCR	Accuracy	Precision	Recall	F1-Score
1	-	-	-	-	88.41	88.27	87.77	88.02
2	✓	-	-	-	88.81	88.75	87.74	88.24
3	✓	✓	-	-	88.41	88.05	87.32	87.68
4	✓	✓	✓	-	89.06	88.68	87.77	88.22
5	✓	✓	✓	✓	90.10	89.78	89.29	89.53

, five different cases were evaluated by progressively enabling the DRRF blocks and the GCCR module to assess their individual and combined impact on classification performance. Case 1 represents the baseline DRC-Net configuration without any DRRF or GCCR blocks, achieving an accuracy of 88.41% and an F1-score of 88.02%. In Case 2, the introduction of a single DRRF block resulted in slight improvements across all metrics, indicating that feature refinement through DRRF contributes positively to the learning process. In Cases 3 and 4, additional DRRF blocks were incrementally added. While the performance gains remained modest, Case 4 showed a noticeable improvement, with accuracy increasing to 89.06%. This suggests that stacking multiple DRRF blocks can enhance hierarchical feature representation when appropriately configured.

The best performance was achieved in Case 5, where three DRRF blocks were combined with the proposed GCCR module. This configuration attained an accuracy of 90.10%, precision of 89.78%, recall of 89.29%, and F1-score of 89.53%. Compared to the baseline DRC-Net, this represents an improvement of 1.69% in accuracy and 1.51% in F1-score. Overall, the ablation results confirm that while DRRF blocks incrementally improve feature discrimination, the integration of the GCCR module is crucial for capturing global contextual relationships. The combination of DRRF and GCCR forms a more robust DRC-Net architecture, leading to consistent performance improvements in heavy construction equipment classification.

3.3. Comparison of HCFF-Net with State-of-the-Art (SOTA) Methods

The effectiveness of the proposed model was evaluated by comparing it with several SOTA approaches. To ensure fairness, consistent experimental protocols were applied, and all methods were evaluated on the same dataset comprising 10 categories. As shown in

Table 5. Performance comparisons of HCFF-Net with different SOTA methods on the ACID.

Models	Accuracy	Precision	Recall	F1-Score
Shuffle-Net [60]	77.21	77.52	74.38	75.92
AlexNet [61]	80.71	80.47	79.78	80.13
InceptionNet-v3 [62]	84.61	82.97	83.60	83.29
VGG16 [41]	84.76	84.20	83.01	83.60
XceptionNet [63]	86.66	86.00	86.00	86.00
EfficientNet-b0 [64]	87.01	86.45	86.18	86.31
MobileNet-v2 [65]	87.81	87.26	87.41	87.33
ResNet50 [66]	88.11	87.69	87.42	87.55
DenseNet121 [53]	88.41	88.18	87.16	87.67
ConvNext-S [52]	88.41	88.27	87.77	88.02
Proposed (HCFF-Net)	90.60	90.60	89.52	90.05

, the proposed HCFF-Net achieves the best overall performance with an accuracy of 90.60%, which is higher than other SOTA methods. This corresponds to an improvement of 2.19% accuracy over ConvNeXt-S and DenseNet121 and 2.49% accuracy over ResNet50. Similar improvements are observed in precision (90.60%), recall (89.52%), and F1-score (90.05%), confirming the effectiveness of the proposed hybrid feature fusion strategy.

While the numerical improvements over existing methods may appear moderate, their significance becomes more evident in the context of real-world construction environments. Construction sites are highly dynamic and visually complex; even small improvements in classification accuracy can substantially enhance the reliability of automated monitoring systems. The proposed HCFF-Net demonstrates improved robustness under challenging conditions such as occlusion, background clutter, and scale variation, which are commonly observed in construction scenarios. Therefore, the performance gains achieved by HCFF-Net are not only statistically meaningful but also practically important for enabling reliable equipment monitoring and data-driven decision-making in construction management systems.

3.4. Comparisons of Model Complexity

Model complexity was evaluated across four key aspects: the number of trainable parameters, floating-point operations (FLOPs), memory usage, and inference time. As shown in

Table 6. Computational complexity comparisons of HCFF-Net with different SOTA methods on the ACID.

Models	Memory Usage (MB)	#Param (M)	FLOPs (G)
Shuffle-Net [60]	1.37	0.35	0.044
AlexNet [61]	217.61	57.04	0.71
InceptionNet-v3 [62]	103.75	23.83	2.86
VGG16 [41]	512.32	134.30	15.47
XceptionNet [63]	79.66	20.83	4.59
EfficientNet-b0 [64]	15.50	4.02	0.41
MobileNet-v2 [65]	8.66	2.24	0.33
ResNet50 [66]	89.96	23.52	4.13
DenseNet121 [53]	26.89	7	2.89
ConvNext-S [52]	188.69	49.45	8.69
Proposed (HCFF-Net)	247.83	64.82	12.77

, HCFF-Net exhibits higher computational complexity, with 64.82 M parameters compared to the baseline models ConvNeXt-S (49.45 M) and DenseNet121 (7 M). However, this increase in complexity is justified by a clear and consistent improvement in classification performance. For instance, compared to ConvNeXt-S, HCFF-Net increases the number of parameters by approximately 31%, while achieving a 2.19% improvement in classification accuracy and a 2.03% gain in F1-score (Table 5). This indicates that the additional parameters are effectively utilized to enhance feature representation and robustness, particularly under challenging conditions. Furthermore, when compared to larger models such as VGG16, which contains 134.30 M parameters, HCFF-Net achieves superior performance (90.60% vs. 84.76% accuracy) while using approximately 52% fewer parameters. This demonstrates that HCFF-Net not only improves performance over competitive baselines but also maintains a more efficient parameter utilization compared to heavier architecture.

Processing time was also evaluated in both GPU and CPU modes, as presented in

Table 7. Inference time comparisons of HCFF-Net with different SOTA methods on the ACID.

Models	Desktop		GPU Mode
-	Inference Time	Frame/s	Inference Time	Frame/s
Shuffle-Net [60]	16.73	59.78	10.58	94.49
AlexNet [61]	27.61	36.21	3.38	295.90
InceptionNet-v3 [62]	113.98	8.77	18.05	55.40
VGG16 [41]	244.34	4.09	20.19	49.52
XceptionNet [63]	126.61	7.93	13.19	75.82
EfficientNet-b0 [64]	42.41	23.58	13.11	76.26
MobileNet-v2 [65]	28.82	34.70	8.26	121.03
ResNet50 [66]	110.31	9.07	11.04	90.58
DenseNet121 [53]	112.24	8.91	25.06	39.90
ConvNext-S [52]	163.27	6.12	23.62	42.34
Proposed (HCFF-Net)	304.85	3.28	43.54	22.97

. The hardware specifications are described in Section 2. As shown in Table 7, HCFF-Net requires higher computation time (43.54 ms GPU inference) compared to lightweight models such as MobileNet-v2 (8.26 ms) and ShuffleNet (10.58 ms). However, these lightweight models achieve significantly lower classification accuracy (87.81% and 77.21%, respectively), as shown in Table 5, indicating that HCFF-Net prioritizes classification robustness over extreme computational efficiency. Compared to ConvNeXt-S, the proposed model improves accuracy by 2.19% (Table 5), while the inference time increases by about 84%, which is expected due to the dual-backbone design and additional feature refinement modules. For real-world applications, the frame rate (FPS) typically needs to be in the range of 20–30 FPS, and the proposed model achieves 22.97 FPS, which is acceptable for deployment on smart edge devices.

3.5. Confusion Matrix Results

A confusion matrix analysis was conducted to examine class-wise misclassification patterns and to reveal the inherent complexities of visually similar construction activities. The confusion matrix provides a summary of the predictions in a matrix form [67], showing both correct and incorrect classifications for each class. In the matrix, the diagonal entries represent $TP$ for each class, while the off-diagonal entries indicate $FP$ and $FN$, respectively. The confusion matrices obtained for the proposed HCFF-Net, on the ACID, are presented in Figure 8.

3.6. Comparison with the Second-Best Model

The performance comparison between the proposed HCFF-Net and the second-best model (baseline model) is illustrated in Figure 9, where HCFF-Net achieves higher performance across all evaluation metrics, including accuracy, precision, recall, and F1-score.

To further analyze class-wise performance differences, a difference confusion matrix is presented in Figure 10. Positive values indicate classes where HCFF-Net achieved higher accuracy, whereas negative values represent classes where the second-best model performed better as computed in Equation (13).

$$D=C^A-C^B$$

(13)

where $C^A$ and $C^B$ represent the confusion matrices of HCFF-Net and the second-best model, respectively. The diagonal elements represent class-wise improvements in correct predictions, while off-diagonal elements indicate changes in misclassification patterns between the two models.

Additionally,

Table 8. Possible outcomes of two models.

No.	Second-Best Model Correct	Second-Best Model Incorrect
HCFF-Net correct	1729	84
HCFF-Net incorrect	40	148

presents the McNemar contingency table used for statistical comparison between the two models. Both models correctly classified 1729 samples. HCFF-Net correctly classified 84 samples that were misclassified by the second-best model (denoted as b), whereas the second-best model correctly classified 40 samples that were misclassified by HCFF-Net (denoted as c). Furthermore, 148 samples were misclassified by both models. McNemar’s test yielded ${\chi }^2=14.91$ with $p=0.00011$, indicating that the performance difference between HCFF-Net and the second-best model is statistically significant ($p<0.05$).

3.7. Visualization Result Using Grad-CAM

To provide a deeper understanding of the model’s decision-making process, the gradient-weighted class activation mapping (Grad-CAM) technique [68] was utilized. Grad-CAM is an effective visualization tool that highlights the regions of an input image that most influence the model’s prediction. In other words, it helps identify where the model is looking when making classification decisions. This is particularly important in construction environments, where images often contain cluttered backgrounds and multiple overlapping objects. The primary objective of this analysis is to verify whether the proposed HCFF-Net focuses on meaningful and discriminative regions of heavy equipment, rather than irrelevant background information. This allows us to interpret the model behavior and validate the effectiveness of the proposed feature fusion strategy.

The correctly classified samples are shown in Figure 11, where the model predominantly focuses on the relevant equipment regions corresponding to the target class. By focusing on global and class-specific characteristics instead of shared components, the model extracted discriminative contextual and structural representations, leading to accurate predictions. The red region in the heat map represents strong activation and high confidence in class prediction, indicating that the model effectively captures the discriminative structural features of each equipment class, such as the vehicle body, operational components, and distinctive shapes. This focused attention on class-specific features enables the model to distinguish between different equipment categories accurately, resulting in correct classification outcomes.

Despite the remarkable overall performance of the proposed method, some heavy equipment classes exhibited classification errors as shown in Figure 12. The primary reason for this misclassification is the high visual similarity between different equipment categories, particularly in terms of shape, color, and texture. In Figure 12a, the dozer class is misclassified as an excavator because the machine is not fully visible. The model mainly focused on the lower track region, which looks very similar to the tracked base of an excavator, leading to an incorrect prediction. In Figure 12b, the target class is dump truck which is misclassified as wheel loader due to the presence of large front structural parts, which leads the model to attend to shared features rather than class-specific characteristics. In Figure 12c, the target class is excavator, but the image is misclassified as dump truck because the model emphasizes the vehicle body and container-like structure while failing to fully capture the distinctive excavator arm features. Similarly, in Figure 12d, the target class is mobile crane, but the image is misclassified as dump truck, as the model focuses on the large vehicle body and ignores the extended crane structure, resulting in incorrect prediction. In Figure 12e, the target class is wheel loader, but the image is misclassified as dump truck because the dominant cabin features are more prominent than the bucket arm details. This confusion is due to feature dominance in the scene, rather than a limitation in the model’s ability to learn robust representations.

4. Discussion

This section discusses the main findings of the study in comparison with previous research on heavy equipment classification. In addition, statistical analysis is provided to demonstrate the performance improvement achieved by the proposed approach compared with the second-best-performing method. Finally, the limitations of the proposed approach are discussed.

There is limited existing research specifically addressing heavy equipment classification in complex construction environments. For example, Nath et al. [41] applied transfer learning using a VGG-16 architecture to construction object classification. Their approach primarily relied on pretrained feature extraction without introducing architectural modifications to enhance feature representation. In contrast, the proposed HCFF-Net introduces dedicated feature enhancement modules, namely, the DRRF and GCCR blocks, which are specifically designed to improve contextual feature learning and robustness under challenging construction site conditions such as occlusion, background clutter, and scale variation. These architectural improvements enable HCFF-Net to learn more discriminative structural and contextual features compared to conventional transfer learning approaches. Another important difference lies in dataset design and evaluation complexity. Nath et al. constructed their dataset using web-mined images collected through limited keyword searches such as truck, dozer, excavator, and crane, which restrict equipment diversity and reduce real-world variability. In contrast, this study evaluates the model using the publicly available ACID benchmark dataset, which contains ten heavy equipment categories collected from real construction environments. This dataset includes diverse viewpoints, background clutter, illumination variations, and occlusion scenarios, providing a more realistic evaluation of classification robustness. The increased number of equipment categories and environmental variability makes the classification task more challenging and better reflects practical construction monitoring conditions.

Similarly, the proposed results were compared with previous CNN-based heavy equipment classification research such as by Yamany et al. [42], who developed a CNN-based model for classifying 12 types of construction equipment and reported approximately 80% precision. Their approach relied on a conventional CNN architecture without incorporating specialized feature enhancement mechanisms for improving contextual understanding. In contrast, HCFF-Net introduces hybrid contextual feature fusion through the integration of DRRF and GCCR modules, which improves both local structural feature extraction and global contextual awareness. Another important methodological difference is the dataset-splitting strategy. Yamany et al. used only 10% of the dataset for testing, whereas the proposed study uses 20% of the dataset for testing. Using a larger test set generally provides a more reliable and unbiased evaluation of model generalization because the performance is validated on a larger number of unseen samples. Furthermore, their study reported the results without comparison with competing methods, whereas the present study on HCFF-Net provides a comprehensive comparison with various state-of-the-art deep learning architectures, as presented in Table 5. This comparison demonstrates the competitive performance of HCFF-Net across multiple evaluation metrics and strengthens the validity of the proposed approach.

Although the proposed HCFF-Net demonstrates strong performance, some failure cases are observed in specific classes. One limitation is found in the dump truck class, where only 65 out of 146 samples were correctly classified, resulting in an error rate of 55.5%, as shown in Figure 8. A detailed analysis of the confusion matrix shows that most errors occur due to confusion with excavators (56 samples, 38.4%) and wheel loaders (23 samples, 15.8%). These quantitative results indicate that dump trucks share visual similarities with these equipment types, particularly in construction environments where trucks are frequently positioned near excavators during loading operations, as shown in Figure 13. This spatial co-occurrence and similarity in structural components contribute to classification ambiguity. This observation is further supported by analysis of the misclassification patterns of the excavator and wheel loader classes. For example, the excavator class shows a comparatively lower error rate of 12.6% (38 out of 301 samples). Most misclassifications occur with dump trucks (20 samples, 6.6%) and wheel loaders (11 samples, 3.6%), typically in cases involving partial visibility of the boom or bucket or overlapping machinery, as shown in Figure 14.

To mitigate these limitations, data-centric strategies can be considered. From the dataset distribution shown in Table 2, the dump truck class contains 729 samples, which is significantly lower than visually similar classes such as excavator (1506 samples) and wheel loader (1324 samples). This imbalance can limit the model’s ability to learn sufficiently discriminative features for the dump truck category, leading to higher misclassification rates. Targeted data augmentation for underrepresented classes such as dump trucks could help balance the training distribution and improve classification robustness. Techniques such as rotation, scaling, and contrast variation could increase the diversity of heavy equipment and improve generalization under complex visual conditions. Moreover, contrastive learning could improve feature separability by pushing confusing classes farther apart in the embedding space while preserving intra-class consistency.

5. Conclusions

This study presented HCFF-Net, a hybrid contextual feature fusion network designed to address the relatively underexplored problem of heavy equipment classification in complex construction environments. To address limitations in existing approaches, this work introduced a specialized deep learning architecture that enhances feature representation through contextual feature fusion and multi-scale feature learning. By integrating the proposed DRRF and GCCR modules, the network effectively captures both local structural characteristics and global contextual relationships, which are essential for distinguishing visually similar construction equipment under realistic site conditions.

The main contributions of this work to the field of AI-based construction monitoring can be summarized as follows:

• A novel deep learning framework (HCFF-Net) was developed specifically for heavy equipment classification in visually complex construction environments.
• Two feature enhancement modules (DRRF and GCCR) were introduced to improve contextual feature representation and classification robustness.
• A hybrid contextual feature fusion strategy was proposed to improve discrimination between visually similar equipment categories.
• A comprehensive evaluation methodology was established, including ablation analysis, statistical validation, and visualization-based model interpretation to ensure reliable performance assessment.

The main findings obtained from the experimental evaluation can be summarized as follows:

• HCFF-Net achieved superior classification performance compared to conventional deep learning architectures, reaching 90.60% accuracy and 90.05% F1-score, demonstrating the effectiveness of contextual feature fusion.
• Ablation experiments confirmed that the DRRF and GCCR modules significantly improve feature discrimination and contribute to overall performance improvement.
• Comparative evaluation showed that HCFF-Net outperforms several state-of-the-art CNN architectures while maintaining competitive computational efficiency.
• Statistical validation using McNemar’s test demonstrated that the performance improvement over the second-best performing model is statistically significant.
• Confusion matrix analysis and Grad-CAM visualization confirmed that HCFF-Net successfully learns discriminative structural features and focuses on relevant equipment regions during classification.

Despite these contributions, some limitations remain that define important directions for future research. First, the model still shows misclassification between visually similar equipment types. Second, this study focuses only on image-based classification and does not incorporate temporal information from videos or multimodal data sources.

Future work can extend this study in several important directions. First, incorporating advanced attention mechanisms may further improve discrimination between visually similar equipment categories. Second, integrating multimodal data sources such as video sequences, sensor data, or operational metadata could further improve robustness. Finally, future research will focus on real-time deployment and edge implementation of the proposed model within intelligent construction monitoring systems to support practical field applications.