Beyond Complexity: Evaluating Simple vs. Complex Deep Learning Architectures for Urban Density Mapping

Abstract

The spatial distribution of population density and land use is a fundamental metric for targeting spatial planning policies under the Sustainable Development Goals. This study presents a comparative evaluation of deep learning architectures for urban density estimation, with an emphasis on spatial generalization. The study was conducted in Jakarta and its surrounding cities (Jabodetabek), a highly heterogeneous region that provides a rigorous setting for evaluating model generalization. Sentinel-2 satellite imagery and European Space Agency WorldCover 2020 data were acquired through Google Earth Engine, together with the Normalized Difference Vegetation Index and the Normalized Difference Built-up Index to explicitly represent vegetation and built-up characteristics. Quantitative evaluation showed that all deep learning models performed well, with R² exceeding 0.94 in all experiments. The U-Net architecture achieved the highest predictive accuracy, with the lowest mean squared error (MSE = 0.0006954). Notably, the lightweight convolutional neural network (CNN) proved to be a highly competitive alternative (MSE = 0.0006998), requiring approximately four times fewer parameters and delivering roughly twice the inference speed of the deeper U-Net and ResNet models. In contrast, complex models such as EfficientNet and ResNet showed diminishing returns in accuracy relative to their design complexity, while classical machine learning baselines (Random Forest and XGBoost) failed to capture spatial configurations robustly. These findings demonstrate that increasing network depth does not necessarily improve performance in spatial fractional regression. Therefore, model selection should balance predictive accuracy against computational and environmental costs.



Received: 10 March 2026 | Revised: 9 June 2026 | Accepted: 3 September 2026



Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.



Data Availability Statement

The data that support the findings of this study are openly available in the Google Earth Engine Data Catalog at https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2_SR_HARMONIZED (Sentinel-2 Level-2A Harmonized), the Google Earth Engine Data Catalog at https://developers.google.com/earth-engine/datasets/catalog/ESA_WorldCover_v100 (ESA WorldCover v100), and Kaggle at https://www.kaggle.com/datasets/nwilhelm123/urban-density-fractional-regression (processed GeoTIFF files).



Author Contribution Statement

Willy Boen: Conceptualization, Methodology, Software, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Visualization. Edy Irwansyah: Conceptualization, Methodology, Formal analysis, Writing – review & editing, Supervision, Project administration.