Towards High-Resolution Population Mapping: Leveraging Open Data, Remote Sensing, and AI for Geospatial Analysis in Developing Country Cities—A Case Study of Bangkok

Abstract

This study develops a globally adaptable and scalable methodology for high-resolution, building-level population mapping, integrating Earth observation techniques, geospatial data acquisition, and machine learning to enhance population estimation in rapidly urbanizing cities, particularly in developing countries. Using Bangkok, Thailand, as a case study, this research presents a problem-driven approach that leverages open geospatial data, including Overture Maps and OpenStreetMap (OSM), alongside Digital Elevation Models, to overcome limitations in data availability, granularity, and quality. This study integrates morphological terrain analysis and machine learning-based classification models to estimate building ancillary attributes such as footprint, height, and usage, applying micro-dasymetric mapping techniques to refine population distribution estimates. The findings reveal a notable degree of accuracy within residential zones, whereas performance in commercial and cultural areas indicates room for improvement. Challenges identified in mixed-use and townhouse building types are attributed to issues of misclassification and constraints in input data. The research underscores the importance of geospatial AI and remote sensing in resolving urban data scarcity challenges. By addressing critical gaps in geospatial data acquisition and processing, this study provides scalable, cost-effective solutions in the integration of multi-source remote sensing data and machine learning that contribute to sustainable urban development, disaster resilience, and resource planning. The findings reinforce the transformative role of open-access geospatial data in Earth observation applications, supporting real-time decision-making and enhanced urban resilience strategies in rapidly evolving environments.

1. Introduction

1.1. Background

Access to population data is essential for comprehensively understanding demographic dynamics, which is crucial for analyzing residents’ needs and planning policies accordingly. While population growth rates in many developed countries have slowed, urbanization continues to accelerate, particularly in the cities of developing countries where population growth remains significant [1,2,3,4,5]. In addition, the United Nations’ initiatives aim to systematically record national populations to monitor progress toward the 2030 Agenda for Sustainable Development [6].

For example, Figure 1 illustrates the trends in urban and rural populations worldwide, in Japan (a developed country), and in Thailand (a developing country and the study area of this study). In Japan, the rural population began to decline rapidly around 2000, while urban population growth also plateaued around 2010. In contrast, in Thailand, the rural population started to decline gradually around 2000, whereas rapid urban population growth began, surpassing the rural population in the late 2010s. Such urban population growth trends are not unique to Thailand but are widely observed in many developing countries worldwide, showing a surpassing urban population share around 2000 [1,7]. This transition has underscored critical urban planning difficulties, especially in rapidly expanding metropolitan areas like Bangkok, where infrastructure deficits, environmental degradation, and social disparities have become increasingly pronounced. These issues, explored in greater detail in the subsequent sections, highlight the urgent need for robust and accurate population mapping to address urbanization challenges effectively.

Figure 1. Urban and rural population trends from 1960 to 2020 in the world, Japan, and Thailand.

1.2. Challenges Posed by Urban Population Growth

Urban population growth imposes significant challenges, including excessive infrastructure burdens, increased environmental impacts, and widening social inequalities. In rapidly urbanizing regions, existing urban planning frameworks and public services often struggle to keep pace with the escalating population, leading to severe issues such as traffic congestion, housing shortages, and unstable water and electricity supply systems [8]. Moreover, higher population densities exacerbate air pollution and waste management issues, further compromising urban sustainability [9]. These challenges not only diminish overall quality of life but also negatively affect economic activities and public health. Additionally, abrupt urban expansion alters land use patterns, intensifying environmental risks such as flooding and land subsidence. Many cities are experiencing reduced green spaces and wetlands due to population-driven urban sprawl, resulting in increased impervious surfaces that degrade natural drainage systems and heighten flood risks [10]. These issues are particularly pronounced in regions vulnerable to climate change, underscoring the urgent need to enhance urban resilience.

The situation is especially critical in developing countries, where rapid urbanization exacerbates existing deficiencies in urban planning, infrastructure, and housing [11]. The shortage of adequate housing has led to the proliferation of informal settlements, where access to fundamental public services such as potable water, sanitation, and electricity remains severely limited [12,13,14]. Consequently, deteriorating public health conditions and entrenched poverty cycles further deepen social inequalities. Additionally, inadequate transportation infrastructure results in prolonged commuting times and higher mobility costs, diminishing economic productivity while also contributing to worsening air pollution [15,16]. From a disaster risk perspective, unregulated urban expansion often leads to the clustering of structurally weak buildings in hazard-prone areas, such as floodplains and landslide-prone slopes, amplifying vulnerability to natural disasters [17,18].

Thus, urban population growth, particularly in developing regions, manifests as a complex interplay of infrastructure deficits, escalating environmental degradation, expanding social disparities, and increased disaster risks. Addressing these multifaceted challenges requires accurate and high-resolution population distribution data, which are essential for effective urban planning and informed policymaking.

1.3. Existing Studies and Challenges in Population Mapping

Understanding urban population distribution is critical for effective urban planning, and population mapping has been widely recognized as an essential tool in this context. Traditional approaches have relied on government-conducted population censuses, leading to the development of choropleth maps, which aggregate population data by administrative units; dasymetric mapping, which redistributes population estimates based on ancillary data; and gridded population datasets, which provide spatially disaggregated estimates [19,20,21]. However, these methods are heavily dependent on census data, which are often updated infrequently and may suffer from accuracy limitations. The challenge is particularly pronounced in developing countries, where access to up-to-date census data is limited, and significant discrepancies between official statistics and actual population distributions have been reported [22,23].

Numerous nations have faced significant challenges in ensuring timely updates to their census data, a situation further complicated by disruptions caused by the COVID-19 pandemic, which have impacted both data collection capabilities and funding [24,25,26,27,28]. Thailand, which serves as the study area for this research, has not conducted a comprehensive update of its national population census since 2010. While these datasets offer a broad overview, their temporal relevance has been compromised due to delays exacerbated by the pandemic, limiting their applicability in dynamic urban analysis [24,25]. In addition to census data, many governments, including Thailand’s, have adopted resident registration systems to maintain population records. However, an evaluation of these datasets within the Bangkok Metropolitan Region revealed notable inconsistencies, indicating that resident registration data do not accurately reflect actual population figures (Figure 2). A comparative analysis between the 2010 national census and concurrent resident registration data demonstrates significant statistical divergence (a 32.3% lower registered population than enumerated, representing approximately 2.8 million individuals, with R² of 0.66) [23]. These discrepancies suggest the presence of extensive informal settlements, a common phenomenon in many rapidly urbanizing regions of the Global South.

Figure 2. Spatial visualization of the population difference between the 2010 household registration records and the 2010 population census in the Bangkok Metropolitan Region, Thailand, at the district level.

The challenges associated with obtaining accurate population data in developing countries stem from multiple interrelated factors. The fundamental dataset for population analysis is typically derived from official population records [21,29,30] but the data collection process is fraught with obstacles, including financial constraints [28,31,32,33], privacy concerns [31,34], limited accessibility [33,35,36], and significant temporal lags between data collection and publication [37]. Such delays significantly undermine the accuracy of demographic analyses and subsequently impair effective policy formulation. For example, India has postponed its national census until 2026 despite projections indicating that its population would surpass China’s within this period. This reliance on outdated demographic information has negatively impacted policy planning and welfare distribution [26,27]. Such temporal discrepancies often result in inaccuracies, as traditional census data may fail to reflect real-time population densities and distributions accurately. In contrast, some high-income regions, such as Japan [38] and EU countries [39], have been developing and maintaining gridded population datasets with finer spatial resolutions, often made available as open data. However, in most other countries, population data continue to be aggregated at coarser administrative levels, such as municipalities or districts, limiting their applicability for detailed spatial analyses. These limitations include temporal lags between censuses that can introduce significant biases, particularly in rapidly urbanizing areas undergoing substantial demographic shifts. Such temporal discrepancies often result in inaccuracies, as traditional census data may fail to reflect real-time population densities and distributions accurately [27].

Furthermore, while some commercially available population datasets exist [40,41,42], cost remains a significant barrier to their widespread adoption in developing countries. The financial burden associated with procuring and maintaining high-resolution, proprietary population datasets poses a substantial challenge, making it infeasible for many governments and research institutions to integrate such data into large-scale, long-term urban planning and management frameworks [43].

1.4. Research Objectives

Taking these challenges into account, this study aims to address the aforementioned challenges, particularly those related to the development and maintenance of high-precision population datasets in developing countries. To achieve this, we propose a cost-effective and globally applicable approach that leverages open data to facilitate the accurate and sustainable collection, update, and analysis of population information. Specifically, we developed a micro-dasymetric population mapping method at the building level, which enables fine-scale population estimation. The dataset generated through this approach is referred to as “Micro Population Data”. By implementing this building-scale population mapping, we enhanced the ability to conduct micro-scale analyses [44], providing valuable insights into community needs, infrastructure demands, and environmental impacts.

To achieve these goals, this study focuses on:

• Developing a globally applicable micro-dasymetric population mapping framework to enable high-resolution, building-level population estimation by incorporating methods for imputing missing attributes.
• Leveraging open geospatial data and remote sensing technologies to enhance the accuracy and scalability of population mapping while integrating advanced data processing techniques to refine missing or incomplete building attributes.
• Testing and validating the model using openly available datasets and integrating advanced population mapping techniques with geospatial tools and machine learning-based imputation methods to enhance estimation accuracy.
• Establishing a cost-effective and continuously updatable population estimation model to overcome the limitations of traditional census-based approaches, ensuring scalability and long-term applicability in data-scarce environments, particularly in developing countries.

Our methodology integrates advanced population mapping techniques with multiple open-source data streams to develop a scalable and cost-effective solution for data-scarce environments. Given the absence of temporally consistent reference data, the effectiveness of this approach will be assessed using multiple population datasets as benchmarks. By leveraging a method initially developed in Japan, where high-quality reference data were available, we aimed to extend and test the model in Bangkok, Thailand. This region exhibits contrasting levels of data reliability, allowing us to evaluate the model’s adaptability across diverse urban contexts. To address potential discrepancies, we incorporated modifications to the framework, ensuring alignment with the characteristics of regions where the original model may encounter limitations.

By integrating innovative machine learning techniques with openly accessible datasets, this research seeks to bridge critical gaps in population mapping. It contributes to sustainable urban planning and policymaking by providing high-resolution, actionable insights into population distribution while ensuring accessibility for resource-constrained communities worldwide.

2. Material

This study aims to evaluate the transferability of Japan’s established urban analysis methodology to Thailand, testing its applicability. Japan represents an optimal methodological baseline due to its exceptionally comprehensive spatial data infrastructure, characterized by fine-resolution building attribute data derived from systematic surveys [45] and population data available at granular mesh sizes of 250 m [46]. In contrast, developing regions typically operate with significantly constrained data environments. This research extends previous frameworks by implementing an analysis in Bangkok using alternative open datasets, thereby investigating methodological adaptability in data-limited contexts.

2.1. Study Area

This study was conducted in Bangkok, the capital of Thailand, located in Central Thailand, which was selected as the study area due to its rapid urbanization and significant population growth. Bangkok is geographically situated along the banks of the Chao Phraya River, covering an area of approximately 1566 square kilometers. The city is home to over 5 million registered residents, with an additional 3 million unregistered migrants [23], collectively accounting for approximately 15% of the country’s total population. With a population density exceeding 6600 people per square kilometer, Bangkok ranks among the most densely populated cities in Southeast Asia. It serves as a political, cultural, and economic hub for the region, experiencing intense urban expansion and significant land use changes over the past few decades. Its combination of rapid urban growth, high population density, and dynamic land use changes presents unique challenges for urban planning, disaster risk management, and environmental sustainability.

Bangkok’s rapid population increase, driven by both natural growth and large-scale rural-to-urban migration, has placed immense pressure on urban infrastructure, public services, and environmental sustainability. The city has experienced widespread land cover transformations, with former agricultural and green areas converted into dense urban developments. These changes have led to increased impervious surfaces, which exacerbate flood risks, a critical issue in low-lying urban regions [47]. Given these dynamics, high-resolution population data and advanced geospatial analysis techniques are essential to understanding the evolving urban landscape and supporting sustainable development planning.

From a geospatial perspective, the study area consists predominantly of a low-lying alluvial plain with relatively flat terrain, as illustrated in Figure 3. Terrain complexity was identified as an important factor influencing spatial variability, particularly in terms of flood susceptibility and urban heat island effects. This complexity was quantified using the Terrain Ruggedness Index (TRI), which measures the standard deviation of elevation differences between a point and its adjacent points. TRI values for the area were derived from NASADEM data with a spatial resolution of 30 m. The analysis yielded a mean TRI value of 4.20 m with a standard deviation of 2.76 m, indicating moderate variability in terrain ruggedness across the study area.

Figure 3. Study area.

2.2. Data Sources

The dataset employed in this study consists of Overture Map Building Footprints, OpenStreetMap (OSM) Point-of-Interest (POI) data, household size and composition, and Digital Elevation Models (DEM), including both the Digital Surface Model (DSM) and the Digital Terrain Model (DTM), as detailed in Table 1. These datasets, derived from Earth observation and open geospatial sources, facilitate the acquisition of high-resolution spatial information, essential for modeling urban population distributions.

Table 1. Information on the source data of this study.

Data Type	Provider	Temporal Coverage	Publication Date	Source
Building footprint	Overture Maps	2024 *	18 December 2024	[48]
POI	OSM	2024 *	N/A	[49]
Aerial images	Google Earth	2024 *	N/A
DSM	ALOS World 3D-30 m (ALOS AW3D30)	March 2016	7 December 2016	[50]
DEM	NASADEM Global Digital Elevation Model	21 February 2000	4 April 2021	[51]
Household size and composition	United Nations	January 2019	2021	[52]

* The dataset is continuously being updated based on the provided information.

To support building height estimation for this study, among many available global DEMs, ALOS AW3D30 and NASADEM were selected based on key criteria such as open access, temporal coverage, spatial resolution, and alignment with this study’s objectives. While advanced DEMs like LiDAR offer unparalleled spatial resolution and accuracy for urban studies [53,54,55,56], their limited global accessibility is highly constrained [56], making them unsuitable for broad-scale applications or scenarios prioritizing open data use in many of the regions. The choice of ALOS AW3D30 and NASADEM is based on the balance between resolution, accuracy, and data accessibility, as detailed in Table 2. ALOS AW3D30 is known for its high accuracy, with a vertical resolution of 5 m, a spatial resolution of 30 m, and global coverage up to March 2016 [50]. NASADEM, in contrast, serves as a DTM and provides terrain data from February 2000, maintaining a comparable spatial resolution [51]. While NASADEM’s temporal coverage is older, it remains relevant due to minimal changes in urban ground surfaces over time. Both models provide global spatial coverage and are widely used in urban and regional planning, each offering distinct advantages in terms of resolution and accuracy. Although NASADEM is nominally a DTM, our analysis revealed the presence of surface objects in the data, not presenting ground elevation exclusively. Therefore, ALOS AW3D30 is used as the primary source for height estimation, and NASADEM is used as a supplement ground reference.

Aerial imagery from Google Earth presents a challenge in balancing accessibility with the precision required for this study. While alternatives may offer more openly licensed data, Google Earth provides high-quality satellite imagery that is valuable for geospatial analysis. As this research is categorized as academic, its use is permissible [57]. However, careful attention must be given to licensing restrictions when replicating this study.

Building footprint data used in this study were sourced from Overture Maps, which extends OSM through machine learning techniques applied by major technology companies [48]. OSM, as a crowdsourced geospatial database, provides critical building information, including footprints, usage, and height or number of floors [49]. However, global data completeness remains a significant limitation, as the overall utilization of OSM building data is estimated at less than 20%, while height-related attributes are available for fewer than 5% of structures [58,59]. Moreover, these attributes are predominantly available in well-maintained regions, restricting their broader applicability [59]. To address these limitations, Overture Maps enhances OSM by supplementing its attributes, thereby making it a more comprehensive resource for urban spatial analysis [48].

Table 2. Detailed information on the selected DEMs [60].

Characteristic	AW3D30	NASADEM
Spatial resolution (m)	30	30
Vertical accuracy (m)	<5	3.5
Datum	ITRF97 and GRS80, using EGM96	WGS84/EGM96
Methodology	Photogrammetry	Interferometric SAR
Data source	ALOS PRISM	SRTM, ASTER GDEM, ICESat

Table 2. Detailed information on the selected DEMs [60].

Characteristic	AW3D30	NASADEM
Spatial resolution (m)	30	30
Vertical accuracy (m)	<5	3.5
Datum	ITRF97 and GRS80, using EGM96	WGS84/EGM96
Methodology	Photogrammetry	Interferometric SAR
Data source	ALOS PRISM	SRTM, ASTER GDEM, ICESat

Household size and composition, as reported by the United Nations, are derived from meticulously maintained census data and global coverage and are available for public access [52].

2.3. Validation and Training Data

The validation data used in this study originates from our field survey conducted to gather comprehensive building information for model evaluation and accuracy assessment. The survey was necessary as there were no public or open-access data available. These data serve as a crucial benchmark for assessing the accuracy of building footprint extraction and height estimation, and the collection methodology and attribute details are outlined in Appendix A. The survey area was selected based on recent urban expansion trends in the Bangkok Metropolitan Region, as reported by the Policy and Planning Division, Office of City Planning and Development, Bangkok [61].

To further enhance validation, high-accuracy building height data were incorporated from NTT Data Corporation, providing a robust reference for evaluating the performance of remote sensing-based height estimations. This dataset is commercially available and considered proprietary, offering high-precision building height information derived from the stereo image processing of satellite data. Given the limited availability of globally consistent LiDAR data, this dataset provides an alternative high-resolution validation source for urban-scale applications.

The geographical coverage of the validation data encompasses the Vadhana District and the Saphan Sung District of Bangkok, representing a diverse urban morphology with varying building densities and land use patterns. The dataset includes approximately 110,000 buildings, offering a significant sample for model evaluation. The height data were extracted from the ALOS AW3D Enhanced DSM, a commercial dataset with a maximum spatial resolution of 0.5 m and a vertical accuracy of 1 m. The integration of high-resolution DEMs and satellite-derived elevation models ensures a rigorous validation framework for population estimation and urban planning applications.

By leveraging multi-source validation data, including our field survey observations and high-resolution DSM-based dataset, this study ensures a robust evaluation framework that aligns with the objectives of remote sensing-based urban analysis. These validation datasets provide key insights into the accuracy of open-source geospatial data and their applicability to scalable population mapping models.

3. Methods

The methodology section begins with an overview of population distribution modeling, followed by the development of attribute data, a critical component in this study. Attribute data serve as essential input variables for the population estimation framework, enabling a more accurate and spatially refined representation of urban population distribution. This study integrates remote sensing-derived datasets, open geospatial data, and advanced spatial analysis techniques to enhance the precision and scalability of the estimation framework. The methodology employs a combination of machine learning-based data imputation, remote sensing data fusion, and high-resolution spatial modeling, ensuring that missing attributes are effectively compensated for while enhancing the reliability and scalability of population estimation in diverse urban environments. This problem-driven approach directly addresses data availability and granularity challenges that hinder traditional census-based methods. The overall workflow of the research is illustrated in Figure 4. It demonstrates how multi-source data are integrated into a structured and reproducible analytical pipeline for urban population mapping.

Figure 4. Workflow of this study.

3.1. Population Distribution

This study builds upon previous research that leverages building information datasets [60,62] for high-resolution population distribution mapping. The key attributes considered include building location, footprint, height, number of floors, usage, and realistic population fluctuations. These attributes play a crucial role in accurately estimating population distribution, particularly in densely populated and rapidly urbanizing regions. Table 3 outlines the different scenarios for population distribution analysis under varying levels of data availability, demonstrating how data completeness influences the accuracy of the population estimation framework. The population distribution algorithm employs a dasymetric pro rata approach, integrating population statistics and household numbers, which are then allocated at the individual building level based on building-specific attributes.

Table 3. Population distribution analysis considering data availability constraints.

Case	Data Availability	Allocation Method
1	Building location only	Population/households are distributed evenly across all buildings in the area, assuming equal importance.
2	Building location and footprint	Allocation is adjusted based on building size, with larger buildings receiving a proportionally higher share of population/households.
3	Building location, footprint, and height/number of floors	Allocation is proportional to building volume (calculated as footprint × height/number of floors).
4	Building location, footprint, and height/number of floors and usage data	Allocation is narrowed to residential buildings only (e.g., detached houses, condominiums), excluding non-residential buildings unless required.
5	Building location, footprint, and height/number of floors; usage data; and realistic fluctuations	Allocation incorporates variability, ensuring that residential units have differing household/resident numbers rather than uniform distribution.

In this study, open data sources such as Overture Maps, which extends to OpenStreetMap (OSM) [63], provided comprehensive building footprint data, ensuring the accuracy of building location information. However, a major limitation of these datasets is the lack of detailed building attributes, particularly height and usage [59]. Since these attributes significantly impact the accuracy of population distribution modeling, further refinement is required. To address this limitation, machine learning-based imputation techniques and classification methods were employed to enhance the completeness of building height and usage data, enabling a more robust and scalable population estimation framework. By integrating remote sensing-derived elevation models and auxiliary geospatial datasets, the availability and accuracy of these essential building attributes can be significantly improved.

The initial step in the population distribution methodology involves determining the volume of each building (v_ij) by multiplying the building’s area (s_ij) by its number of floors (f_ij), expressed as follows:

$$v_{ij}=s_{ij}\times f_{ij}$$

(1)

where

• v_ij: volume of the building j in subarea i.
• s_ij: area of the building j in subarea i.
• f_ij: number of floors of building j in subarea i.

In cases where the floor number is not specified, the building height was utilized to estimate the number of floors. This estimation follows the guidance provided by the Building Control Act, B.E. 2522 [64] and assumes an average floor height of 3.0 m, consistent with the study area standards.

Subsequently, the number of households was allocated to each building based on its relative volume within the subarea. The number of households assigned to a building (h_ij) is calculated as follows:

$$h_{ij}=H_i\left({\displaystyle \frac{v_{ij}}{\sum _{k=1}^m{v}_{ik}}}\right)$$

(2)

where

• h_ij: number of households assigned to building j in subarea i.
• H_i: total number of households in subarea i.
• m: number of buildings in the subarea.

To refine the population distribution estimates further, the estimated area of each household $\left(h{s}_{ij}\right)$ is derived from the building volume and the number of households allocated to that building for each building category. The estimated household area is calculated as:

$$h{s}_{ij}={\displaystyle \frac{v_{ij}}{h_{ij}}}$$

(3)

where ${hs}_{ij}$ is the estimated area of each household in building j in subarea i.

Finally, the number of residents allocated to each household (r_ij) is determined based on the estimated household area. This is expressed as:

$$r_{ij}=R_i\left({\displaystyle \frac{h{s}_{ij}}{\sum _{k=1}^{m}h{s}_{ik}}}\right)$$

(4)

where

• r_ij: number of residents assigned to building j in subarea i.
• R_i: total number of residents in subarea i.

This methodological approach allows for an accurate estimation of population distribution by accounting for the relative volumes of buildings and the corresponding household areas within each subarea. This approach improves the reliability of population allocation by incorporating variations in building sizes, household distributions, and subarea characteristics.

3.2. Building Attributes

3.2.1. Building Height Estimation

Ensuring reproducibility and accessibility, this study utilized open-access remote sensing datasets to estimate building heights. Specifically, photogrammetry-derived Digital Surface Models (DSM) were employed to extract building heights, while Interferometric Synthetic Aperture Radar (InSAR)-derived Digital Elevation Models (DEM) served as ground reference data. The methodology was adapted to leverage the globally available and openly accessible ALOS AW3D30 dataset [60,65].

There have been numerous studies on estimating building heights using machine learning-based approaches [66,67], as well as research focusing on the utilization of remote sensing-derived products [65,68,69]. Additionally, several investigations have employed high-resolution DEMs and point cloud data, including InSAR and LiDAR techniques [54,56,70,71,72]. While these methods have demonstrated high accuracy, their applicability is often constrained by data accessibility, high computational costs, and the lack of global coverage. Current state-of-the-art research often employs high-resolution elevation data (up to 10 m) to produce lower-resolution building height maps (90 m) [69,72] or covers extensive US areas using LiDAR [56], indicating the demand for high-quality input for the estimation. In contrast, this study prioritizes the use of open-source datasets, allowing for a scalable, cost-effective, and globally applicable approach. While this may result in lower data quality compared to proprietary datasets, a comprehensive evaluation of its performance is conducted to ensure the feasibility and reliability of open-access remote sensing data for urban analysis.

The core principle of the methodology involves calculating the difference between ground and non-ground elevation points [53,55]. This study employs a DSM and a Digital Terrain Model (DTM) to derive the Surface Height Model (SHM), which represents the height of surface objects:

$${\mathrm{f}}_{\mathrm{S}\mathrm{H}\mathrm{M}}={\mathrm{f}}_{\mathrm{A}\mathrm{W}3\mathrm{D}30}-{\mathrm{f}}_{\mathrm{D}\mathrm{T}\mathrm{M}}$$

(5)

where

• f_SHM: surface height model.
• f_AW3D30: surface elevation from AW3D30, non-ground points.
• f_DTM: terrain elevation derives from AW3D30, ground points.

The DTM, used to represent the terrain for calculating surface object height, was derived from the ALOS AW3D30 dataset via morphological erosion, which can be expressed as:

$${\mathrm{f}}_{\mathrm{D}\mathrm{T}\mathrm{M}}={\epsilon }_B\left({\mathrm{f}}_{\mathrm{A}\mathrm{W}3\mathrm{D}30}\right)$$

(6)

where

• ε_B(∙): morphological erosion operation with structuring element B.
• B: structuring element (SE).

To account for terrain variations, a terrain correction was then performed, particularly for slopes exceeding 10 percent [65], which were identified using Horn’s slope algorithms [73]. The correction was based on NASADEM as the reference DTM. In terrain areas with steep slopes, the overestimation of building heights may occur. The height correction for a slope (${\mathrm{s}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e}}_{{\mathrm{c}\mathrm{o}\mathrm{r}}_{ij}}$) is determining by computing the differences between the eroded ${\mathrm{f}}_{\mathrm{D}\mathrm{T}\mathrm{M}}$, which represents the “bottom of the slope (BOS)”, and the dilated ${\mathrm{f}}_{\mathrm{D}\mathrm{T}\mathrm{M}}$, which represents the “top of the slope (TOS)”. The corrected surface height model $cSH{M}_{ij}$ at a given location is expressed as:

$$\mathrm{c}\mathrm{S}\mathrm{H}{\mathrm{M}}_{\mathrm{i}\mathrm{j}}=\left\{\begin{array}{c}S{HM}_{ij}-{\mathrm{s}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e}}_{{\mathrm{c}\mathrm{o}\mathrm{r}}_{ij}}, \mathrm{i}\mathrm{f} {\mathrm{s}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e}}_{ij} \ge x\%\\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}S{HM}_{ij}, \mathrm{i}\mathrm{f} {\mathrm{s}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e}}_{ij}<x\%\end{array}\right.$$

(7)

where

• cSHM_ij: corrected SHM at position (i,j).
• SHM_ij: SHM at position (i,j).
• slope_ij: slope value at position (i,j).
• $Slop{e}_{co{r}_{ij}}$: slope correction based on terrain.
• x: threshold slope percentage.

Building height estimates were computed as gross height $H_{gross}$, and net usable height $H_{net}$ used the following equations:

$$H_{gross}\left({BF}_i\right)=\max \{cSHM\left(x,y\right)\mid \left(x,y\right)\in {BF}_i\}$$

(8)

$$H_{net}\left({BF}_i\right)={\displaystyle \frac{1}{N}}\sum _{\left(x,y\right)\in B_i}cSHM\left(x,y\right)$$

(9)

where

• SHM(x,y): elevation value at a given pixel location (x,y).
• BF_i: building footprint i.
• N: total number of pixels within the building footprint.

By integrating open-access DSM and DTM datasets, combined with robust spatial modeling techniques, this study ensured that building height estimation remained scalable, cost-effective, and globally applicable. Despite potential data quality limitations compared to proprietary datasets, the methodology emphasizes repeatability, adaptability, and validation to enhance the usability of open-source remote sensing data in urban analysis.

3.2.2. Building Use Classification

Several approaches for urban building use classification methods exist in the literature, utilizing aerial photographs, street-view imagery, and graph-based methods [58,74,75]. While these methods have demonstrated considerable success, they are often limited by data accessibility, computational requirements, and scalability issues. Despite advances in these technologies, balancing data quality, cost, and global applicability remains a critical consideration for practical implementation in urban analysis. To address these challenges, this study prioritizes the use of open-source and widely accessible datasets, ensuring that the methodology remains scalable, cost-effective, and globally reproducible, particularly for rapidly urbanizing regions with limited data availability.

This study implements a dual-modal classification framework that integrates geospatial feature analysis and computer vision techniques to enhance the accuracy and scalability of urban building use classification. Given the diversity of the available data modalities, two distinct model architectures were developed [60,76], as visualized in Figure 5:

Figure 5. Building use classification workflow.

• Polygon-Based Classification Model.
• Image-Based Classification Model.

An ensemble approach was applied to derive building use classifications, combining polygon-based and image-based classification models through an arithmetic averaging of their predicted probabilities. This strategy minimizes individual bias and enhances overall classification robustness.

The Polygon-Based Classification Model processes tabular data derived from building polygons and urban context features using LightGBM (Light Gradient Boosting Machine), as depicted in Figure 6. The features relevant to classification include geometric attributes (area, perimeter-to-area ratio, footprint complexity), proximity metrics to POIs, and road network characteristics [76]. A detailed description of these attributes is provided in Appendix B (Table A7 and Table A8).

Figure 6. Polygon-based classification scheme.

The selection of LightGBM as the classification model was guided by the consideration of model capabilities. Traditional linear classifiers (e.g., logistic regression) are inadequate for modeling complex nonlinear spatial relationships inherent in urban datasets [77]. Although alternative nonlinear methods like Random Forests and Support Vector Machines (SVM) are effective at modeling complex interactions, they often benefit from preprocessing steps to address missing data. Given that urban datasets, particularly POI and road network data, often contain missing values at broader spatial scales, it was essential to select a model capable of inherently managing incomplete records.

Compared to other gradient-boosting decision tree (GBDT) implementations like XGBoost or CatBoost, LightGBM offers the native handling of missing values, computational efficiency through histogram-based algorithms, and an asymmetric tree-growth strategy, making it especially suitable for large-scale urban analytics [78]. Thus, LightGBM aligns closely with this study’s goals of scalability, accuracy, cost-effectiveness, and global reproducibility.

The Image-Based Classification Model (Figure 7) was implemented using ResNet-50, a Convolutional Neural Network (CNN) architecture, with transfer learning initialization. This model processes aerial imagery validated through field-collected ground truth data, enhancing classification accuracy by leveraging deep feature representations extracted from this high-resolution aerial imagery, making it an effective complement to polygon-based classification. This model’s ability to analyze spatial texture and visual patterns extends the framework’s precision in urban contexts, particularly where geospatial tabular features alone may not provide sufficient granularity.

Figure 7. Image-based classification scheme.

By integrating these two complementary classification models and utilizing remote sensing and geospatial data, this study establishes a highly scalable, cost-efficient, and globally applicable building use classification framework. Unlike existing methodologies that require high-cost proprietary data or computationally intensive deep learning models [74,75], this approach maintains a balance between accuracy, affordability, and accessibility, making it particularly valuable for developing regions experiencing rapid urban expansion.

4. Results

4.1. Building Height Estimation Results

An experiment was undertaken to approximate building heights using an open data source, which encompassed building footprints from the Overture Map. The primary Digital Surface Model (DSM) dataset utilized was ALOS AW3D30. The visualization is presented in Figure 8.

Figure 8. Visualization of the estimated building height in Bangkok (over the Pathum Wan and Vadhana Districts).

The estimation process is performed using erosion techniques on Digital Surface Model (DSM) data. It was verified that the DSM pixel size substantially surpasses the building size. However, the building patch is deemed significant in relation to both the building size and the pixel size. The footprint of detached houses relative to the pixel size is illustrated in Figure 9. By examining the average footprint size of buildings, the optimal size for the structuring element is established.

Figure 9. A 30 m (equivalent to 1 pixel of DEMs used) mesh over the building footprint.

Bangkok’s relatively flat geographical characteristics suggest that smoother terrain facilitates better estimation of building heights. Previous studies indicate that building height estimation is more accurate in less complex terrains, thus providing a positive expectation for our results. Table 4 presents a detailed evaluation of determining the optimal size of the structuring element (B) utilized for detecting surface objects, ranging from 3 × 3 pixels (equivalent to 90 m side-length) to 33 × 33 pixels (990 m).

Table 4. Building height estimation accuracy assessment.

Structuring Element Dimension (B)	R2	MAE (m)	RMSE (m)	Accuracy of 5 m Confidence (%)
3 × 3	−0.122	6.26	10.79	58.12
5 × 5	0.002	5.18	9.99	71.90
7 × 7	0.035	4.68	9.71	77.44
9 × 9	0.052	4.39	9.55	80.18
11 × 11	0.060	4.20	9.46	81.79
13 × 13	0.062	4.09	9.40	82.54
15 × 15	0.061	4.01	9.37	82.94
17 × 17	0.058	3.96	9.35	83.13
19 × 19	0.053	3.93	9.34	83.21
21 × 21	0.045	3.91	9.34	83.07
23 × 23	0.036	3.92	9.36	82.81
25 × 25	0.022	3.95	9.40	82.31
27 × 27	0.006	4.00	9.44	81.73
29 × 29	−0.015	4.07	9.51	81.06
31 × 31	−0.035	4.15	9.58	80.46
33 × 33	−0.051	4.22	9.63	79.75
Best		3.91	9.34	83.21

Considering the size of the structuring element that yields the lowest MAE and RMSE and the highest accuracy, an accuracy of 5 m was used in the assessment, as it aligns with the quality of the input DSM data. The 19 × 19 structuring element was found to perform best. It is important that the structuring member be relatively large compared to the building, allowing us to infer that the general building patch is approximately 570 m.

Despite satisfactory MAE and RMSE values, the coefficient of determination (R²) remains relatively low, consistent with findings in similar geospatial applications. Morphological operations at our operational resolution inevitably homogenize complex building structures toward local means. This aligns with our observation that very high building structures may not be well captured, resulting in smoothed-out height details. This explains the reduced predictive variance while maintaining acceptable absolute errors.

This resolution-induced constraint could potentially be mitigated by scaling the resolution or employing modifiers to improve the accuracy of building block estimation. Based on our evaluation, we selected the structuring element B of 19 × 19 for subsequent height estimation phases.

4.2. Building Use Classification Results

The classification results are presented in Table 5 and Table 6, providing a structured evaluation of the building use classification model. The classification categories analyzed in this study include townhouses, detached houses, mixed-use buildings, and others, derived from training data obtained through field surveys.

Table 5. Performance metrics for building use classification.

Building Use	Precision	Recall	F1-Score
Townhouse	0.646	0.316	0.424
Detached house	0.777	0.905	0.836
Mixed-use building	0.716	0.653	0.683
Others	0.791	0.447	0.571
Overall Accuracy	0.755

Table 6. Building use classification normalized confusion matrix.

Ground Truth	Classified Number of Structures
	Townhouse	Detached House	Mixed-Use Buildings	Others	Total Count
Townhouse	0.3156	0.4889	0.1867	0.0089	225
Detached house	0.0218	0.9054	0.0704	0.0025	1194
Mixed-use building	0.0180	0.3219	0.6529	0.0072	556
Others	0.0395	0.2763	0.2368	0.4474	76
Total count	110	1391	507	43	2051

The model achieved the highest performance for detached houses, with a precision of 0.777, recall of 0.905, and F1-score of 0.836. Subsequently, its performance was validated with a corrected classification rate of 90.54% considering that this category constitutes the largest portion of the dataset at 58.21% (1194 out of 2051). This demonstrates the model’s robust capability to manage this class effectively.

For mixed-use buildings, the model achieved moderate performance metrics (precision: 0.716, recall: 0.653, F1-score: 0.683). The classification model demonstrated promising results by correctly identifying 65.29% of buildings within this specific class. This performance is particularly noteworthy because the building class itself possesses inherent ambiguities that challenge even human observation. The model’s capability to accurately recognize approximately two-thirds of instances within this ambiguous building category suggests significant potential for this classification approach.

Furthermore, the “others” category showed high precision (0.791) but lower recall (0.447), resulting in an F1-score of 0.571. The classification model was able to identify buildings of this class correctly 44.74% of the time. This type of building is often situated amidst structures of other kinds, which are also challenging for human observation. This performance indicates the model’s capability to recognize non-conventional building uses within common structures.

Notable classification challenges are evident with townhouses, which exhibited the lowest performance metrics (precision: 0.646, recall: 0.316, F1-score: 0.424), and only 31.56% of townhouses were correctly classified. This outcome prompted a thorough investigation into the factors contributing to these misclassifications. For instance, 48.89% of buildings were categorized as detached houses and 18.67% as mixed-use buildings, highlighting the complexity of this category.

A plausible explanation is the variability in the quality of the building footprint input data, which may have influenced the classification accuracy. This insight provides a promising opportunity for refining data quality and enhancing the overall performance of the classification model.

Overall, the model achieved an overall accuracy of 0.755, indicating that it performs strongly in certain categories, particularly for detached houses, and demonstrates a promising ability to classify mixed-use buildings. While some misclassifications occur between similar categories, such as townhouses and detached houses, as well as mixed-use buildings and detached houses, these findings highlight valuable areas for further refinement. Future improvements in feature selection, model optimization, or the integration of additional data sources may enhance classification accuracy and further refine the model’s ability to distinguish between complex building types.

4.3. Population Estimation Results

After determining building height and usage, unit ratios were calculated using survey data on building usage and household characteristics to aid in population distribution. Initially, the floor area required for a single household in each building type is established based on our survey information (see Appendix A). These unit ratios represent the attributes of the surveyed area and serve as benchmarks for estimating populations at the building level. In Bangkok, the floor area per household differs across various building types:

• One household per unit in detached houses.
• A total of 16.16 m² per household in mixed-use buildings.
• A total of 78.66 m² per household in townhouses.

Household numbers were then assigned to residential buildings by type, with the method varying based on building usage due to differences in household distribution. For instance, mixed-use buildings usually have one household per floor, whereas townhouses allocate households along the building’s length, with the floor count generally not being related to household numbers. Adjustments were made to account for these differences.

Subsequently, a population was assigned to each household probabilistically using data from the UN’s 2019 Household Size and Composition for Thailand [52], shown in Table 7. Specifically, household sizes of two to three and four to five people are allocated based on probabilities, with sizes selected randomly within these ranges. For households of six or more, a fixed value of six people is assigned. This method generates building-level population statistics for Bangkok, as shown in Figure 10.

Table 7. Household size and composition (excerpted).

Country (Year of Aggregation)	Number of People in the Household (%)
	1	2–3	4–5	6+
Thailand (2019)	21.50	48.62	23.39	6.41

Figure 10. Building population estimate in Vadhana district.

The accuracy assessment of the population estimation was conducted by comparing this study’s estimates with various authoritative sources, including census data, registration records, and existing gridded datasets (Table 8 and Table 9). Notably, the household count by the registration record was not disclosed, indicating a gap in the available validation data sources. An aggregate analysis of the population count for Bangkok reveals significant variance compared to the existing population data.

Table 8. Estimated population and population data references.

Dataset Name	Year	Total Population	Difference (%)	Source
Population estimation	2024 *	10,093,488	-	This study
2010 Population Census	2010	8,294,235	21.69%	[24]
2010 Registration Record	2010	5,611,918	79.86%	[79]
High-Resolution Settlement Layer	2015	9,210,179	9.59%	[80]
WorldPop 2015	2015	6,963,596	44.95%	[81]
Global Human Settlement	2019	9,273,267	8.85%	[30]
2021 Registration Record	2021	5,440,544	85.52%	[79]

* The temporal coverage is detailed in Table 1.

Table 9. Household estimation and household data references.

Dataset Name	Year	Total Household	Difference (%)	Source
Household estimation	2024 *	3,515,175	-	This study
2010 Population Census	2010	2,881,752	21.98%	[24]

* The temporal coverage is detailed in Table 1.

Specifically, our analysis using planning zones (zones 1–6; see Table 10) uncovered meaningful patterns regarding estimation precision. Urban centers, particularly the Cultural Conservation area (zone 1) and the Central Business and Commercial District (zone 2), exhibited lower population figures compared to the known census and settlement data, indicating spatial methodological challenges in dense, dynamic, inner-city environments.

Table 10. Bangkok’s six planning zones [61].

Zone Number	Zone Name in English
1	Cultural Conservation and Tourism Promotion Area
2	Central Business and Commercial District
3	Residential Area
4	Suburban Residential and Agricultural Area (Eastern)
5	Suburban Residential and Agricultural Area (Northwestern)
6	Suburban Residential and Agricultural Area (Southwestern)

However, in suburban and residential contexts (notably zone 3, as depicted in Figure 11 and the peri-center area in Figure 12), our population estimations demonstrated notably higher accuracy. Significantly, urban–suburban transitional zones displayed the greatest correspondence, likely due to clearly identifiable settlement typologies (e.g., detached houses), which facilitated accurate classification (confirmed by observed high classification accuracy in Table 5 and Table 6).

Figure 11. Zonal share of the population comparison.

Figure 12. A map of estimated population compared to (a) 2010 population census and (b) Global Human Settlement Layer.

These spatial patterns convincingly align with the documented urban development processes, specifically the ongoing suburbanization characterized by outward residential expansion, alongside complementary inner-city gentrification trends. Consequently, the observed population distribution patterns provide robust empirical support for current urbanization theories and underscore the model’s practical applicability for urban planning and socioeconomic forecasting.

Additionally, comparative validation against the Global Human Settlement dataset (Figure 13) revealed similar spatial correlation patterns and moderate correlation coefficients with the authoritative population census (0.48–0.50) when contextualized against a correlation of approximately 0.66 observed between the official population records and authoritative census data [23]. Such congruence enhances methodological confidence, particularly given that Global Human Settlement data synthesize census-derived and alternative data sources [30,81].

Figure 13. Estimated population compared to (a) 2010 population census and (b) Global Human Settlement Layer.

Overall, despite inherent methodological complexities identified with population enumeration in highly dynamic urban settings, these findings provide valuable confirmation of both the robustness of the estimation approach and its relevance to understanding contemporary urbanization and gentrification dynamics.

5. Discussion

This discussion is structured into three key aspects: the morphological approach for building height estimation, machine learning-based building use classification, and population distribution analysis. These components collectively demonstrate the scalability and applicability of our methodology, particularly in the context of rapidly urbanizing regions where high-resolution data are often limited.

5.1. Morphological Approach for Estimating Building Height Estimation

The morphological erosion approach using ALOS AW3D30 demonstrated moderate accuracy (MAE: 3.91 m, RMSE: 9.34 m) but faced inherent limitations associated with low-resolution (30 m) DEMs (Figure 9). Previous studies have applied morphological operations on neighboring pixels using a 3 × 3 structuring element (covering 90 m²) [65]. However, our findings indicate that such structuring elements are too large to precisely identify individual buildings yet are too small to effectively detect clustered structures, particularly in urban environments where buildings are densely arranged in patches. These results align with previous research [82,83], reinforcing that structuring element size significantly impacts the performance of morphological approaches.

While the 19 × 19 structuring element optimized the performance in Bangkok’s predominantly flat terrain, a dynamic structuring element size and higher-resolution DEMs (e.g., <5 m) would be more effective in resolving mixed-use high-rises in central business districts such as Vadhana. Furthermore, our results suggest a systematic underestimation of building height at the pixel level, particularly in high-density areas. Future implementations could prioritize multi-scale morphological operators to mitigate the resolution-to-footprint mismatch and further enhance building height estimation accuracy.

Challenges remain, particularly regarding the limitations of current open geospatial data and computational workflows. The application of morphological operations to open-source DEM data presents certain inaccuracies in building height estimation, primarily due to resolution constraints. The selection of structuring elements in morphological analysis was found to be highly sensitive to terrain complexity, necessitating further refinement for applications in diverse urban landscapes.

5.2. Machine Learning for Building Use Classification

The findings underscore critical challenges in urban environments, particularly where mixed-use developments and informal spatial arrangements [84] pose significant difficulties for conventional building use classification frameworks. This emphasizes the importance of adaptability to spatially diverse environments. Previous research [58,75] has demonstrated the potential of integrating contextual data such as proximity to points of interest (POIs) and street-level imagery to improve classification performance, especially in distinguishing between commercial and residential structures. However, such methods, including advanced approaches like graph neural networks for facade analysis, are computationally intense, posing a scalability challenge in cities with a large volume of unstructured data. This underscores the necessity of employing locally tuned training data, particularly in regions with weak zoning enforcement.

This study achieved 90.5% classification accuracy for detached houses, yet performance was notably lower for townhouses (31.6%) and mixed-use buildings (65.3%). A key factor of misclassifications includes errors in footprint segmentation from open datasets where terraced structures in Overture Maps were frequently mis-segmented as merged units, as shown in Figure 14b, a problem that was not observed in datasets derived from proprietary sources such as ALOS-derived footprints, as shown in Figure 14a.

Figure 14. Building footprint comparison between (a) proprietary dataset, (b) ML-segmented open dataset, and (c) aerial image.

By integrating the estimated building height with the classified building uses, we can further utilize the data. It is also important to consider the temporal appropriateness of the building footprint, especially in regions experiencing rapid development [44].

5.3. Analysis of Population Distribution

This study advanced the population distribution analysis using a hybrid methodological framework, integrating data from multiple geospatial population layers with household composition data to mitigate census data limitations common in rapidly urbanizing regions. By prioritizing openly accessible datasets, the methodology balances high-resolution output with global scalability, offering a replicable model for data-scarce environments.

Implementing a high-resolution population distribution analysis presents significant challenges regarding data completeness [36], accessibility, and computational feasibility [38]. Our methodology addresses these constraints by leveraging openly available datasets, enabling a viable solution, and thereby providing cost-effectiveness and globally applicable alternatives that reduce dependence on infrequently updated census data.

The methodology’s geographic scalability and temporal consistency are the primary constraints where sampling biases may limit generalizability across urban forms, while infrequent population dataset updates hinder real-time validation. These limitations intertwine with rapid socioeconomic shifts, and Bangkok’s transit-driven gentrification and upscale urban development have demonstrably displaced lower-income populations [85,86,87], creating dynamic demographic patterns that challenge conventional estimation frameworks.

While theoretical consistency between our analysis and the observed urban processes provides provisional validation, definitive confirmation requires future census data to resolve spatiotemporal and socioeconomic complexities. Nevertheless, the framework synergistically addresses conventional census limitations through open-access data integration and advanced spatial modeling. Future work should explore the integration of multi-source remote sensing data and machine learning-based approaches to refine the accuracy and applicability of population distribution estimations in rapidly urbanizing contexts.

6. Conclusions

This study demonstrates the effectiveness and future potential of integrating open geospatial data (crowdsourced building footprints and satellite-derived geomorphology) with machine learning to enhance localized population estimation. By leveraging scalable, cost-effective, and globally accessible methodologies, this research provides valuable insights for urban analytics and spatial planning, particularly in data-scarce environments.

Our approach enables more granular population distribution mapping than spatial resolution-dependent grid. While validation remains challenging due to outdated reference data, the performance was notably robust in urban–suburban transitional zones, which aligns with areas that are overrepresented in validation datasets. The generalizability across various urban morphologies should be heavily taken into consideration for machine learning applications.

Our integrated approach combining morphological erosion with machine learning classification achieved moderate building height estimation accuracy (MAE: 3.91 m, RMSE: 9.34 m) and heterogeneous building use classification performance (90.5% for detached houses, 31.6% for townhouses). Nonetheless, this provides a foundation for developing adaptive and scalable methodologies capable of addressing complex urban spatial patterns. While challenges related to data availability, methodological trade-offs, and local contextual variability persist, this research emphasizes the importance of fostering greater accessibility to high-quality, up-to-date geospatial and demographic data.

Future advancements in this field will require not only technical refinements, such as higher-resolution DEMs and enhanced classification frameworks but also the continuous evolution of Earth observation-based methodologies and geospatial artificial intelligence (AI) for improved population modeling. By incorporating automated data acquisition techniques, cloud computing, and geospatial AI, future research can further refine problem-driven solutions that support sustainable urban development and disaster resilience strategies. Additionally, strengthening interdisciplinary collaboration between geospatial scientists, urban planners, and policymakers will be crucial in harnessing the full potential of Earth observation data and analytical tools. By addressing these challenges, this research advances the broader goal “Towards High-Resolution Population Mapping in Developing Country Cities”, contributing to a more data-driven, sustainable, and resilient approach to urban development worldwide.

Appendix A.1. Validation Data Collection

The data collection was based on the building footprint from the Overture Map. In the field documentation stage, buildings are classified by their appearance characteristics, such as detached houses, detached offices, commercial buildings, condominiums, and mixed-use structures [64], detailed in Table A1 and Figure A2. Additionally, the number of households is identified by address and number of visible mailboxes to imply the number of households; building conditions are assessed with categories like clean, intermediate, and deteriorated; and building height is recorded by the number of floors. The survey sites encompass urban, transitional urban–suburban, and suburban areas, as shown in Table A2, featuring detached housing, townhouses, and high-rise residential neighborhoods. Figure A3 and Figure A4 highlight the buildings informing the informal utilization of the area. This is evident in the random emergence of building uses among other types, which presents one of the challenges considered in this study. Furthermore, the survey in progress is depicted in Figure A5.

Table A1. Survey attributes.

Attribute	Description
Building Use Classification	Commercial, condominium, detached commercial, detached residential, mixed-use (primarily commercial), mixed-use (primarily residential), other, townhouse
Building Condition	Clean, intermediate, or deteriorated
Building Height	The height of the building, expressed as the number of floors
Number of Posts	The count of visible utility or structural posts associated with the building or property
Vacant House Indicator	Identifying whether the property is vacant or unoccupied
X, Y Coordinates	The longitude–latitude value of the building’s geographic location

Table A1. Survey attributes.

Attribute	Description
Building Use Classification	Commercial, condominium, detached commercial, detached residential, mixed-use (primarily commercial), mixed-use (primarily residential), other, townhouse
Building Condition	Clean, intermediate, or deteriorated
Building Height	The height of the building, expressed as the number of floors
Number of Posts	The count of visible utility or structural posts associated with the building or property
Vacant House Indicator	Identifying whether the property is vacant or unoccupied
X, Y Coordinates	The longitude–latitude value of the building’s geographic location

Table A2. Validation area characteristics.

Area	Land Use Characteristics	Characteristics
1. Phaya Thai District	Central Business and Commercial Area	The district is notable for its condominium and commercial development, attracting young workforces. The west side boasts businesses, mid-to-high-end townhouses, shops, and detached buildings. In contrast, the east side features terraced houses and detached buildings.
2. Bang Khen District	Residential Area	Transitional urban–suburban residential area located near the military base and airport. It features smaller enterprises in the townhouses and detached residential units.
3. Bang Kapi District	Residential Area	Located on the eastern side of Bangkok. It boasts a prominent shopping center that serves as a transportation hub, connecting the canal and rail networks. The area also features townhouses along the primary road and large detached residences.
4. Vadhana District	Central Business and Commercial Area	Prominent central business and commercial area. Mix of office buildings, condominiums, and mid- and high-end residential buildings. Features vibrant retail, dining, and entertainment hubs, making it a hotspot for both residents and visitors.
5. Saphan Sung District	Residential Area	Quiet residential area on the east side of the Bangkok city center, comprising well-planned detached houses, townhouses, and gated communities.

Table A2. Validation area characteristics.

Area	Land Use Characteristics	Characteristics
1. Phaya Thai District	Central Business and Commercial Area	The district is notable for its condominium and commercial development, attracting young workforces. The west side boasts businesses, mid-to-high-end townhouses, shops, and detached buildings. In contrast, the east side features terraced houses and detached buildings.
2. Bang Khen District	Residential Area	Transitional urban–suburban residential area located near the military base and airport. It features smaller enterprises in the townhouses and detached residential units.
3. Bang Kapi District	Residential Area	Located on the eastern side of Bangkok. It boasts a prominent shopping center that serves as a transportation hub, connecting the canal and rail networks. The area also features townhouses along the primary road and large detached residences.
4. Vadhana District	Central Business and Commercial Area	Prominent central business and commercial area. Mix of office buildings, condominiums, and mid- and high-end residential buildings. Features vibrant retail, dining, and entertainment hubs, making it a hotspot for both residents and visitors.
5. Saphan Sung District	Residential Area	Quiet residential area on the east side of the Bangkok city center, comprising well-planned detached houses, townhouses, and gated communities.

Figure A1. A screenshot of the software used during the field survey.

Figure A2. Validating data survey sites.

Figure A3. Picture of townhouses in Bang Khen (area 2) being used as commercial buildings surrounded by residential buildings.

Figure A4. Picture of a detached house (in the red box) in Bang Khen (area 2) being used as a commercial building surrounded by residential buildings.

Figure A5. The field survey in progress.

Appendix A.2. Validation Data Sample and Statistics

The data collection process involved a field survey conducted over 6 days from 2023 to 2024. A survey team consisting of four to five members utilized the ArcGIS Field Map for on-site data collection via mobile devices. The methodology employed direct observation and digital documentation of urban structures and their functional uses.

The challenges in surveying emerged during the data collection process, primarily related to the accurate identification of building structures and their functional classifications. These limitations stemmed from the complexity of urban typologies and the multifunctional nature of certain structures, which presented classification challenges even for human observers. Such challenges highlight the inherent difficulties in developing automated classification systems for complex urban environments. Table A3, Table A4, Table A5 and Table A6 show the surveying records and the summary of the results.

Table A3. Surveyed data samples.

ID	Surveyed Flag	Building Use Classification	Building Height	Number of Posts	x	y
1	Surveyed	Detached residential	2	1	100.630983	13.7789598
2	Surveyed	Condominium	3	4	100.617464	13.8918068
3	Surveyed	Mixed-use (primarily residential)	3	2	100.605937	13.8933387
4	Surveyed	Detached residential	2	1	100.627159	13.7745783
5	Surveyed	Detached residential	1	1	100.626576	13.7747678
…	…	…	…	…	…	…

Table A3. Surveyed data samples.

ID	Surveyed Flag	Building Use Classification	Building Height	Number of Posts	x	y
1	Surveyed	Detached residential	2	1	100.630983	13.7789598
2	Surveyed	Condominium	3	4	100.617464	13.8918068
3	Surveyed	Mixed-use (primarily residential)	3	2	100.605937	13.8933387
4	Surveyed	Detached residential	2	1	100.627159	13.7745783
5	Surveyed	Detached residential	1	1	100.626576	13.7747678
…	…	…	…	…	…	…

Table A4. Surveyed building count by district.

	District	Count
1	Phaya Thai	362
2	Bang Khen	424
3	Bang Kapi	636
4	Vadhana	675
5	Saphan Sung	767

Table A4. Surveyed building count by district.

	District	Count
1	Phaya Thai	362
2	Bang Khen	424
3	Bang Kapi	636
4	Vadhana	675
5	Saphan Sung	767

Table A5. Surveyed building count by type.

Building Type	Count
Detached residential	1416
Detached commercial	71
Condominium	277
Mixed-use (primarily residential)	186
Mixed-use (primarily commercial)	199
Commercial	329
Townhouse	101
Others	285
Total	2864

Table A5. Surveyed building count by type.

Building Type	Count
Detached residential	1416
Detached commercial	71
Condominium	277
Mixed-use (primarily residential)	186
Mixed-use (primarily commercial)	199
Commercial	329
Townhouse	101
Others	285
Total	2864

Table A6. Surveyed building count by floor group.

Building Floor	Count
1–3	2344
4–6	419
7–10	55
11-	46

Table A6. Surveyed building count by floor group.

Building Floor	Count
1–3	2344
4–6	419
7–10	55
11-	46