<p>Summary of data sources used in this study.</p>

Accurate traffic volume prediction is essential for managing congestion, improving road safety, mitigating environmental impacts, and supporting long-term transportation planning. The traditional four-step travel demand model (FSM) is a well-established framework, but it relies on static survey data, substantial calibration effort, and simplified behavioural assumptions that may not adequately capture complex travel patterns. In contrast, data-driven models are capable of learning nonlinear relationships from large datasets, yet they are often designed for short-term forecasting and typically do not target the long-term, segment-level volume estimation tasks required for strategic planning. This study proposes Mukara, a deep learning framework that directly approximates the mapping from external socioeconomic and network features to observed traffic volumes on highway trunk road segments. The model is trained on eight years of data from England and Wales and incorporates population, employment, land use, road network characteristics, and points of interest as inputs. Mukara achieves a mean GEH of 50.74, a mean absolute error of 8,989 vehicles per day, and an R2 of 0.583 under random cross-validation, outperforming baseline models and existing studies under comparable settings. Under a more stringent region-based spatial cross-validation scheme, performance remains robust, demonstrating strong spatial transferability. Ablation experiments further demonstrate the robustness of the proposed architecture and reveal the relative importance of different input feature groups for prediction.