To guide the review process, the following research questions were formulated:

RQ1: Which convolutional neural network architectures have demonstrated greater effectiveness in the spatiotemporal prediction of vehicular traffic?

RQ2: What are the main methodological limitations and technical challenges identified in the use of CNNs for traffic prediction?

RQ3: What evaluation metrics and experimental protocols are commonly used to validate CNN models in traffic prediction?

RQ4: How are convolutional neural networks used to model spatial dependencies in traffic flow prediction?

RQ5: How do CNN-based models integrate temporal information to improve accuracy in short- and long-term traffic prediction?

RQ6: What strategies do CNN models employ to incorporate exogenous variables such as weather, events, and accidents into traffic prediction?

These questions were formulated with the aim of delimiting the scope of the study and structuring the review according to the key components of the research problem.

Table 1. Definition of general concepts with PIPOH.

The literature search was conducted in five specialized databases: Redalyc, Scopus, IEEE Xplore, Research Rabbit, and DOAJ. To ensure broad coverage of the scientific literature, search terms related to convolutional neural networks, vehicular traffic prediction, spatiotemporal analysis, deep learning, and hybrid models were used.

The search strategy included combinations of keywords, synonyms, and Boolean operators to retrieve relevant studies published between 2021 and 2026. The search was conducted in fields such as title, abstract, and keywords, prioritizing scientific publications directly related to the study's topic.

Table 2. Terms, synonyms, and filters to compose the search string

Table 3. Main terms and databases.

Table 4. Search string.

The documents retrieved from the databases underwent a 2-stage selection process. First, duplicate records were removed. Subsequently, full-text availability was reviewed, and finally, the previously defined inclusion and exclusion criteria were applied.

Table 5. Dates of consultation and file downloads.

Inclusion criteria:

Studies related to the application of convolutional neural networks in vehicular traffic prediction.

Research focused on spatiotemporal prediction or vehicular flow analysis.

Scientific articles, reviews, conference papers, or book chapters published between 2021 and 2026.

Documents available in full text in Spanish or English.

Exclusion criteria:

Works that did not directly address the use of CNN in vehicular traffic.

Studies focused exclusively on other techniques unrelated to the project's main focus.

Duplicate publications.

Documents without full-text access or with insufficient information for analysis.

Works that fall outside the established time range.

Once the studies were selected, a quality evaluation was applied with the aim of assessing the relevance and methodological soundness of each document. At this stage, criteria such as the clarity of the research objective, the relevance of the problem addressed, the adequacy of the method used, the robustness of the analysis of results, and the consistency between the findings and conclusions were considered. This procedure made it possible to prioritize the most robust studies and reduce the impact of works with low methodological relevance, thus strengthening the confidence of the patient.

bility of the final synthesis.

Data extraction was performed using an analysis matrix that recorded aspects such as author, year, database, objective, type of CNN model, variables used, evaluation metrics, main results, and reported limitations. This organization enabled systematic comparison of the studies and identification of common patterns in the literature.

The information was synthesized qualitatively by grouping the findings into thematic categories related to CNN architectures, temporal integration, exogenous variables, performance metrics, and methodological challenges. This approach provided an overview of the state of the art and helped answer the research questions posed at the beginning of the study.

RQ1: Which CNN-based architectures (pure or hybrid) have demonstrated greater effectiveness in the spatiotemporal prediction of vehicular traffic?

RQ2: What are the main methodological limitations and technical challenges identified in the use of CNN for traffic prediction?

RQ3: What evaluation metrics and experimental protocols are commonly used to validate CNN models in traffic prediction?

RQ4: How are Convolutional Neural Networks (CNN) used to model spatial dependencies in traffic flow prediction?

RQ5: How do CNN-based models integrate temporal information to improve accuracy in short- and long-term traffic prediction?

RQ6: What strategies do CNN models use to incorporate exogenous variables (weather, events, and accidents) into traffic prediction?

Table 6. Quality assessment in the search and selection processes.

Table 7. Studies included in the systematic review.

Various architectures based on convolutional neural networks (CNN) have achieved high performance in the spatiotemporal prediction of vehicular traffic, ranging from purely convolutional models to sophisticated hybrid architectures that integrate CNN with recurrent networks, advanced attention mechanisms, and traditional machine learning techniques.

Hybrid architectures are notable in spatiotemporal traffic prediction: PPMCC-RF-CNN-BiGRU employs CNN for complex spatial features and BiGRU for bidirectional temporal dependencies, reducing RMSE by up to 39% [51]; DATC-STP (CNN Transformer-style) achieves lower MAE and higher SSIM while reducing parameters by 85% on TaxiBJ [33]; and CNN-LightGBM outperforms CNN-LSTM/Transformer with 57.58% lower RMSE in highway congestion [27].

Hybrid CNN-BiLSTM architectures with attention mechanisms are notable for predicting the risk and severity of road accidents, achieving accuracies close to 95%. The CNN captures relevant spatial traffic patterns, BiLSTM models bidirectional temporal dependencies, and the attention module identifies the most influential predictive variables in complex scenarios [26, 45].

The CNN-PSO (Particle Swarm Optimization) combination increases predictive accuracy from 76.5% to 92.0% in smart city environments. Similarly, IA-CSTM uses CNN to model spatial interactions between nearby vehicles and LSTM for temporal evolution, incorporating spatiotemporal attention for dynamic scenarios [44, 5].

Hierarchical CNN-LSTM networks analyze traffic violations by processing spatial sequences using CNN and capturing temporal dependencies between driving events using LSTM, consistently outperforming non-hierarchical methods in predictive accuracy [53].

Comparative studies confirm the superiority of CNNs over traditional RNNs not only in accuracy but also in energy and computational efficiency, particularly in edge computing, where resource optimization is a critical factor [47].

CNNs face fundamental structural limitations when processing inherently non-Euclidean transportation networks. The required transformation into matrix structures distorts actual spatial relationships and the complex connectivity that characterizes urban road networks [35, 52].

The local nature of convolutional kernels severely limits the capture of long-range spatial dependencies, which are essential for modeling congestion propagation between distant urban regions and for understanding the global dynamics of the road system [5, 51].

The conversion of traffic data into matrices/images causes irreversible loss of critical topological information, including the direction of vehicular flow, the hierarchy among road types, and fundamental structural relationships. At the same time, pure CNNs exhibit significant weaknesses in modeling complex long-term temporal dependencies [35, 24, 31].

Static convolutional kernels, once trained, remain fixed during inference, limiting adaptation to dynamic traffic with rapid changes due to incidents, special events, or seasonal variations in mobility. Unbalanced datasets systematically skew towards free-flowing conditions over real congestion [52, 5].

Deep CNN architectures demand prohibitive computational and memory resources for real-time implementation, making them particularly infeasible on edge devices with energy constraints and critical latency for road monitoring systems [35, 6].

Training exhibits slow convergence, high susceptibility to overfitting when datasets are limited, and critical dependence on architectural design, filter size, and hyperparameter selection, requiring exhaustive and costly exploration of the parameter space [27, 2].

The fundamental mismatch between Euclidean CNN grids and the irregular topology of road networks generates systematic distortions. The traditional hybrid CNN-RNN architecture separates spatial and temporal modules, hindering the deep fusion of spatiotemporal information [24, 34, 56].

Mean absolute error (MAE) quantifies direct discrepancies in the native units of the phenomenon, facilitating practical interpretation. Root mean squared error (RMSE) disproportionately penalizes large deviations, which is crucial in scenarios where extreme errors severely impact operational decisions. Mean absolute percentage error (MAPE) allows relative comparisons between heterogeneous datasets [56, 24, 50].

The mean squared error (MSE) constitutes the primary loss function during training. The coefficient of determination (R2) evaluates the proportion of variability explained by the model. The Huber loss offers a robust balance between the sensitivity of MSE and the stability of MAE in the presence of outliers [4, 46].

Classification tasks (congestion and incidents) use precision, recall, accuracy, and F1-score to evaluate the correct identification of events. Computer vision employs intersection over union (IoU) to assess the spatial quality of vehicle/object detection [26, 12].

The evaluation of computational efficiency includes latency, inference time, FLOPs, and model parameters, which are critical indicators for real-time deployment viability and for edge devices with hardware constraints [47, 10, 28].

Experimental protocols employ chronological splitting (70/10/20, 60/20/20, and 80/20) to avoid future contamination. Min-Max/Z-score normalization stabilizes training. Sliding windows convert time series into multi-step supervised formats [35, 44, 5].

Multi-horizon evaluation (15/30/60 minutes) analyzes degradation in predictive performance across horizons. Filters eliminate low-activity regions, thereby reducing noise. The Adam optimizer with early stopping prevents overfitting. Ablation studies validate the contribution of components against traditional baselines [31, 51, 33, 50, 34].

Convolutional Neural Networks (CNNs) have been widely used in traffic prediction due to their ability to model spatial dependencies by automatically extracting local features. In this context, one of the most common strategies is to transform traffic data, such as flow, speed or occupancy, into matrix representations that can be interpreted as images. In these representations, each cell corresponds to a specific geographical location or sensor within the road network, which allows the capture of spatial proximity relationships between different segments [51, 27, 47, 35]. This approach facilitates the application of convolutional operations on structures that reflect the traffic topology, although simplified in a regular grid.

The spatial learning process relies on convolutional filters or kernels that slide over these matrices to identify local patterns and correlations between adjacent regions. These filters enable the modeling of phenomena such as congestion propagation or influence between neighboring segments, for example, by considering the relationship between upstream and downstream sections on a road [27, 24]. Unlike traditional methods, the parameters of these filters are learned automatically through backpropagation, enabling the capture of complex nonlinear relationships in traffic data [4, 22]. Moreover, stacking multiple convolutional layers enables the construction of hierarchical representations, where initial layers capture local patterns and deeper layers model longer-range spatial dependencies [52, 47, 24].

In more complex scenarios, such as vehicle trajectory prediction or interaction among multiple agents, CNNs are used to process structured spatial representations such as grids or social tensors. These structures enable explicit modeling of relationships between vehicles, considering relative distances and interaction patterns within the traffic environment [5, 18]. Complementarily, some architectures incorporate additional mechanisms such as pooling to reduce dimensionality and improve robustness to noise, although in certain cases these operations are omitted to preserve the topological integrity of the road network [35].

Since traffic is an inherently dynamic phenomenon, CNNs are often integrated into hybrid architectures where they act as spatial encoders. In these models, features extracted by convolutional layers are subsequently processed by recurrent networks, such as LSTM or GRU, or by attention mechanisms to capture temporal dependencies [51, 24, 26]. Likewise, recent research has proposed more advanced variants, such as Transformer-inspired CNNs, which use larger kernels to expand the receptive field and capture long-range spatial dependencies without incurring the high computational cost of models based entirely on attention [33].

The forced Euclidean representation on irregular road topology requires zero-padding of sensor-free regions, introducing systematic artifacts and computational inefficiencies. This fundamental limitation motivates the development of GNNs while preserving the natural graph connectivity of road infrastructures [50].

The integration of temporal information into convolutional neural network-based models for traffic prediction is usually carried out through hybrid architectures in which the CNN is responsible for extracting local spatial patterns and a recurrent network, such as LSTM or GRU, models the temporal evolution of the sequence. This scheme allows capturing short- and long-term dependencies in traffic behavior, and has shown good performance in multi-horizon forecasting scenarios [43, 22, 24].

Another frequent strategy consists of using tridimensional convolutions or spatiotemporal blocks that simultaneously process the spatial and temporal dimensions. Instead of treating time as a separate variable, these architectures learn joint correlations between consecutive observations, which facilitates detecting complex dynamics of the road network and improving the model's sensitivity to abrupt flow changes [33, 31, 43].

Complementarily, some approaches organize the time series into proximity, periodicity, and trend components, enabling the model to leverage both recent patterns and longer-range recurring cycles [34, 15, 25].

It has also become common to combine CNNs with attention mechanisms and Transformers to enhance temporal representation. In such cases, the CNN learns local features, while the attention module assigns greater weight to the most relevant historical intervals, reducing the impact of noise and improving the capture of long-range dependencies [42, 49, 1].

In more specialized architectures, temporal integration can be addressed through ConvLSTM, in which the convolution operation is incorporated directly within the recurrent unit. This preserves the spatial structure while updating temporal memory, making it especially useful for data organized as grids or traffic maps [15, 31, 54].

When the goal is to avoid classical recurrence, some approaches rely on dilated temporal convolutions or dual-attention designs that broaden the receptive field without substantially increasing computational cost [50, 29, 35].

External variables are converted into numerical vectors or matrices and incorporated into the CNN input tensor. For example, additional channels in the traffic image represent rain or accidents using Boolean logic [22]. Exogenous variables such as temperature or day type are organized into matrices or tensors and treated as additional channels by the CNN to extract hidden connections [51, 58, 55].

Exogenous data are processed using FC-NN and fused with feature maps from the traffic sensor CNN [23, 35]. For accidents, third-order tensors integrate duration and severity with flow and speed, enabling joint learning of road degradation [40]. CNNs extract context from external factors, while LSTM or BiGRU processes temporal sequences, capturing future impacts such as congestion caused by accidents [51, 52, 45, 29].

Specific schemes encode temporal anomalies, such as distinct representations for weekdays or holidays [42]. Weight factors in hybrid architectures balance structural information from CNN and external dynamics to enable rapid responses to closures [41]. Random Forest or PPMCC select relevant factors to avoid noise and optimize precision [51, 26, 45]. In the absence of real climate data, meta-parameters such as date implicitly capture seasonal patterns [58].

Spatiotemporal attention layers assign dynamic weights to exogenous variables, prioritizing them during disruptive events [45, 26, 52, 5]. External data are integrated into 2D/3D inputs, as in ER-GAN-CNN, to correlate traffic and weather [4]. Models such as MF-CNN or TFFNet use separate branches for real, historical, and external data, fusing them in dense layers; embeddings are concatenated in graphs [34, 50]. MDN with convolutional layers predicts probability distributions under high uncertainty [30].

Fuzzy logic in CNN robustly handles accident uncertainty [3]. Attention modules or Transformers in CNN-SKIPGRU-Transformer dynamically weight exogenous variables [24, 1]. ST-ResNet uses a branch for exogenous variables such as weather, with 1-hot encoding fused residually [57]. In DMVST-Net and STDN, exogenous variables are concatenated with spatial representations before the temporal LSTM [7, 8]. ST-MetaNet employs POIs as meta-knowledge for dynamic weights [20]; DGCNN decomposes tensors for anomalies by adjusting graphs [19]. Attention prioritizes exogenous variables at critical moments [16, 8].