Multi-Source Corrected Nighttime Light Index for Urban Mapping in Small and Medium-Sized Cities

Abstract

Monitoring urban expansion in small and medium-sized cities is essential for assessing sustainable development. Nighttime light (NTL) data, widely available and consistently capturing human activity intensity, provide a useful proxy for delineating built-up areas. However, accurately mapping urban extents from NTL remains challenging due to radiance blooming effect and limited spatial resolution. This study introduces the vegetation–water–building–thermal nighttime urban index (VWBTNUI), a multi-dimensional fusion framework that integrates NTL with spectral and thermal information to reduce spillover effects and enhance physical consistency in urban extent mapping. Using three representative inland cities in Sichuan Province, western China, VWBTNUI was compared with raw NTL data and four widely used composite indices. Results demonstrate that VWBTNUI consistently outperforms existing approaches, achieving overall accuracy (OA) values above 0.88, F1 scores above 0.80 and Kappa coefficients exceeding 0.72. Furthermore, the urban area estimates derived by VWBTNUI maintained relative area errors below 10%. The extracted urban extents also exhibit strong agreement with benchmark products. Relying on globally accessible datasets and simple pixel-level operations, VWBTNUI offers a scalable and reproducible solution for urban monitoring in data-scarce regions. By offering reliable baseline information for regional planning, the approach supports evidence-based governance and contributes to advancing Sustainable Development Goal (SDG) 11 on inclusive, safe, resilient, and sustainable cities.

1. Introduction

Urbanization has been one of the most significant drivers of socioeconomic transformation in China over the last three decades [1,2]. As functional hubs, urban areas centralize innovation, essential services, and specialized facilities, thereby facilitating the flow of capital and labor across the urban–rural continuum [3]. This process is characterized not only by the spatial expansion of built-up land but also by the intensified concentration of population, economic activities, and infrastructure. Given that these socioeconomic functions are physically manifested through human activities and built environments [4], the accurate delineation of urban boundaries is not merely a cartographic exercise but a prerequisite for evaluating the efficiency of urban space in accommodating and organizing such activities. However, most existing urban boundary delineation frameworks have been primarily developed and validated for large coastal metropolises [5,6]. While recent studies have extended urban mapping to inland cities [7,8,9,10], the spatial characterization of small and medium-sized inland cities remains comparatively limited. These cities often face challenges such as fragmented urban morphology and complex geomorphic settings. Despite these constraints, they play a crucial role in regional connectivity and spatial organization, serving as important nodes in broader socio-economic and infrastructural networks [11]. Their heterogeneous built environments and fragile ecosystems make them particularly vulnerable to unplanned sprawl, which risks diluting urban functional intensity and undermines SDGs related to land use efficiency [12,13,14,15,16]. Consequently, there is an urgent need for robust and reproducible monitoring approaches capable of capturing urban expansion patterns and supporting sustainable spatial planning in inland contexts.

Traditional urban extent approaches based on field surveys or census statistics are often labor-intensive, costly, and temporally inconsistent [17,18]. Satellite remote sensing, particularly NTL imagery, provides a practical alternative for such monitoring due to its strong correlation with human activities [19,20,21]. Products such as the Defense Meteorological Satellite Program/Operational Linescan System (DMSP/OLS) and the National Polar-Orbiting Partnership-Visible Infrared Imaging Radiometer Suite (NPP/VIIRS) have enabled numerous studies on global and regional urbanization [7,22,23,24]. Despite these advantages, the utility of NTL for fine-scale boundary delineation is hindered by the coarse spatial resolution and persistent blooming effects [10,25,26], particularly in inland cities where low-density development at the urban fringe becomes indistinguishable from background noise.

To mitigate the limitations of NTL-based mapping, recent efforts have combined NTL with auxiliary spectral or thermal indicators such as the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Water Index (NDWI), the Normalized Difference Built-up Index (NDBI), and land surface temperature (LST) to characterize urban surfaces and suppress background noise [27,28,29]. However, the efficacy of these individual auxiliary variables is often constrained by spectral confusion and environmental noise. Specifically, the spectral similarity between impervious surfaces and barren land often leads NDVI to conflate bare soil with built-up areas [30]. Similarly, NDBI exhibits pronounced instability in peri-urban transitional zones, where heterogeneous mixtures of impervious surfaces, exposed soil, and sparse vegetation produce overlapping spectral signatures [31,32]. Although NDWI is generally effective for water body detection, its performance is often compromised in urban environments by topographic shading and building-induced shadows [33]. Furthermore, the inclusion of LST introduces temporal uncertainty, as its values are highly sensitive to seasonal shifts and atmospheric variability [34,35].

In response to these localized misclassifications, composite indices such as the vegetation–temperature light index (VTLI) and the normalized urban area composite index (NUACI) have been developed to leverage multiple data sources, effectively reducing saturation effects and enhancing intra-urban contrast [36,37]. Despite their improvements, these frameworks predominantly rely on simplified pairwise combinations, which lack the multidimensional robustness required for complex inland environments. In such regions, the high heterogeneity inherent in mixed industrial–residential patterns and fragmented construction zones creates conflated signal signatures, which may limit the effectiveness of low-dimensional indices [38,39,40]. Additionally, model complexity and interpretability present notable challenges. Some indices, exemplified by the human settlement index (HSI) [41], employ intricate formulations that hinder transparent interpretation and reduce their practicality for large-scale or operational applications [19].

Beyond physical indicators, socioeconomic datasets such as point of interest (POI) data and road networks have been introduced to capture functional patterns of human activity [25,42,43]. While valuable in data-rich metropolitan regions, these datasets often suffer from spatial bias, inconsistent coverage, and substantial preprocessing requirements, limiting their applicability in data-constrained inland cities [44,45]. Synthesizing these challenges, specifically the limited diversity of spectral variables, the structural opacity of existing indices, and the accessibility barriers of ancillary data, it becomes evident that current NTL-based methods struggle to account for the pronounced land-surface heterogeneity of small and medium-sized inland cities. Consequently, these limitations lead to inconsistent or unreliable delineation of urban boundaries in geographically diverse contexts.

To bridge these gaps, this study introduces a robust composite index. By synergistically integrating NDVI, the Modified NDWI (MNDWI) [33], NDBI, and LST with NTL data, VWBTNUI is designed to suppress NTL blooming effects and better characterize urban features. Developed using globally accessible datasets, VWBTNUI prioritizes simplicity, scalability, and applicability to underrepresented inland cities. The specific objectives of this study are to: (1) formulate the VWBTNUI architecture by optimizing the integration of multi-source physical indicators; (2) validate the index’s robustness across three representative small- and medium-sized cities in western China; and (3) provide methodological insights for advancing urban boundary extraction in support of sustainable development and spatial planning in data-limited environments.

2. Materials and Methods

2.1. Study Area

Urbanization in China’s western provinces is characterized by fragmented, heterogeneous, and topographically constrained development patterns, yet these regions remain disproportionately underrepresented in existing research [46,47]. Given these distinct spatial attributes and the prevailing research gaps, Sichuan Province offers a representative case for examining urbanization dynamics in inland contexts. Situated within the transition zone between Southwest and Central China, Sichuan serves as a major economic and industrial engine in the western region, a key transportation and communication hub, and a strategic inland gateway linking the region with South and Southeast Asia. Beyond its geopolitical significance, the province’s diverse topography, substantial demographic pressures, and pronounced regional economic disparities create a distinctive environment for analyzing urban expansion. According to the national urban classification [48] and the Seventh National Census data, Sichuan’s urban system exhibits a hierarchical pyramid structure, comprising one megacity, three large cities, eleven medium-sized cities, and twenty-one small cities [49].

To evaluate the robustness of the proposed method under contrasting socioeconomic and geographic conditions, three representative prefecture-level cities of Sichuan were selected: Ziyang (ZY) and Bazhong (BZ), both classified as small cities, and Luzhou (LZ), a medium-sized city (Figure 1;

Table 1. Basic characteristics of the three selected cities (2020).

City	Area (km2)	Resident Population (Million)	GDP (Billion CNY)	Urbanization Rate (%)	Administrative Units
ZY	5747	2.31	80.75	41.29	Yanjiang District, Lezhi County, Anyue County
BZ	12,301	2.71	76.70	46.16	Bazhou District, Enyang District, Tongjiang County, Nanjiang County, Pingchang County
LZ	12,234	4.26	215.72	50.24	Longmatan District, Jiangyang District, Naxi District, Lu County, Hejiang County, Xuyong County, Gulin County

). This selection was guided by four criteria: urban scale, dominant topography, economic function, and proximity to major metropolitan or transportation corridors. Specifically, ZY represents a small city influenced by metropolitan spillover from the provincial capital, Chengdu. BZ exemplifies a mountainous city constrained by topography and limited industrial bases. In contrast, LZ serves as a medium-sized city positioned along the Yangtze River Economic Belt with relatively balanced development. These three cases reflect divergent urban trajectories, providing a rigorous basis for assessing the applicability of the VWBTNUI method.

In this study, the urban area is defined as the functional built-up land within official administrative boundaries, encompassing industrial zones, university campuses, and spatially detached development clusters [50]. Major water bodies, such as large rivers, were excluded to maintain consistency in defining land-based urban fabric. Only the urban districts of the three cities were analyzed, while peripheral counties were omitted to focus on areas experiencing concentrated and active urbanization.

2.2. Data Sources and Preprocessing

A suite of multi-source geospatial datasets was integrated to support index construction, urban boundary extraction, and independent validation (

Table 2. Description of research datasets.

Data Type	Description	Source
VIIRS data	VNL V2 annual product, 2020	Payne Institute for Public Policy at the Colorado School of Mines (https://eogdata.mines.edu/products/vnl/#annual_v2, accessed on 5 July 2025)
Landsat data	All available images in 2020; indices: NDVI, NDWI, MNDWI, NDBI; LST via SMW algorithm	Google Earth Engine (GEE) platform (https://earthengine.google.com/, accessed on 1 July 2025)
Google Earth images	Historical images (ZY: 28 March 2020; BZ: 1 June 2020; LZ: 20 June 2020)	Google Earth
Comparison products	Dataset of Urban Built-up Area in China (1992–2020) V1.0	National Tibetan Plateau/Third Pole Environment Data Center (https://doi.org/10.11888/HumanNat.tpdc.272851, accessed on 20 July 2025)
	Dataset of Built-up Areas of Chinese Cities in 2020	Science Data Bank (http://doi.org/10.11922/sciencedb.j00001.00332, accessed on 25 July 2025)
Administrative boundary	2020 version	National Earth System Science Data Center (http://www.geodata.cn, accessed on 1 July 2025)

). These datasets comprise: (1) the VIIRS Nighttime Light (VNL) annual composite, which served as the primary indicator of anthropogenic illumination; (2) Landsat 8 Operational Land Imager/Thermal Infrared Sensor (OLI/TIRS) data, utilized to derive spectral indices and LST; (3) high-resolution Google Earth imagery, providing a visual reference for sampling and boundary delineation; and (4) two benchmark urban extent products, adopted for comparative performance analysis. Additionally, administrative boundary data were employed to define the spatial extent of the study areas. The core datasets correspond to the year 2020, selected to maximize data availability and cross-source consistency. This choice is partly motivated by the fact that Dataset of Built-up Areas of Chinese Cities in 2020 provides the most complete and consistent coverage for this year. The associated operational parameters and preprocessing thresholds are summarized in

Table A1. Summary of key operational parameters and thresholds in the workflow.

Stage	Operation	Key Parameters/Thresholds
Data Preprocessing (GEE)	Spatial Alignment	VNL resampled to 30 m
	Outlier Mitigation	Winsorization at upper and lower 0.5% quantiles
	Image Selection	Landsat 8 cloud cover filter < 30%
	Cloud Masking	Pixel-level masking via QA_PIXEL bitmask
	Index Compositing	MVC for index compositing
Index Construction (GEE)	LST Normalization	Min-max scaling to [0, 1] range
Urban Extraction	K-means Clustering	Number of clusters K = 2 (Urban vs. Non-urban)
	Algorithm Stability	Random_state = 42; n_init = 10
Post-classification	Polygon Aggregation	Distance: ZY (120 m), BZ (30 m), LZ (60 m)
	Internal Hole Filling	Area threshold: ZY (5.3 km2), BZ (2.0 km2), LZ (20.0 km2)
	Fragmentation Removal	Removal of isolated patches < 5400 m2
Accuracy Assessment	Sample Size	250 stratified random points per city
	Sampling Stratification	Urban-to-non-urban ratio approximately 1:2

.

2.2.1. NTL Data

The VIIRS product offers significant improvements over the earlier DMSP-OLS series, including enhanced radiometric sensitivity and calibration, and reduced saturation in bright urban centers [51]. This study utilized the annual VNL V2 data, which suppresses noise from lunar illumination, stray light, and clouds through a 12-month median filter process and refined quality masks [52]. This dataset, originally at a spatial resolution of 15 arc seconds (~500 m), was resampled to 30 m to ensure spatial alignment with Landsat-derived indices. Subsequently, the data were reprojected to an Albers Equal Area Conic projection and clipped to the administrative boundaries of the study areas. To reduce the influence of outlier radiance values, the data were winsorized at the upper and lower 0.5% quantiles.

2.2.2. Landsat Data

The Landsat program offers long-term, moderate-resolution observations with consistent spectral and radiometric quality, making it a cornerstone dataset for land studies [29,53,54]. For this research, Landsat 8 OLI Level-2 Surface Reflectance and TIRS Brightness Temperature data for the year 2020 were obtained from the GEE platform (Collection 2, Tier1). The GEE platform was selected for its high-performance accessibility and its capacity for large-scale, cloud-based geospatial processing [55].

The data selection and cleaning process involved several rigorous steps. First, all scenes with less than 30% cloud cover were filtered. Subsequently, pixel-level cloud and shadow masking was executed using the QA_PIXEL bitmask to ensure the integrity of the surface reflectance values. To characterize urban built-up features, we computed three key spectral indices: NDVI, MNDWI, and NDBI, based on their standard formulations (Equations (1)–(3)). Each index was generated using the maximum value composite (MVC) approach to capture the optimal seasonal signal for each land-cover type [56].

$$NDVI={\displaystyle \frac{NIR-Red}{NIR+Red}}$$

(1)

$$MNDWI={\displaystyle \frac{Green-SWIR}{Green+SWIR}}$$

(2)

$$NDBI={\displaystyle \frac{SWIR-NIR}{SWIR+NIR}}$$

(3)

where $NIR$, $Red$, $Green$, and $SWIR$ denote the near-infrared, red, green, and short-wave infrared bands, respectively.

LST was retrieved using the Statistical Mono Window (SMW) algorithm, which establishes an empirical relationship between top-of-atmosphere (TOA) brightness temperature and surface temperature through linear regression [57]. This method is widely utilized in remote sensing research due to its computational efficiency and high practical utility [58]. In this study, LST was retrieved using a GEE-based implementation developed by Ermida et al. [59]. The resulting LST composite was also winsorized at the upper and lower 0.5% quantiles and then normalized using Equation (4) to ensure comparability with other indices.

$$NLS{T}_i={\displaystyle \frac{LS{T}_i-LS{T}_{min}}{LS{T}_{max}-LS{T}_{min}}}$$

(4)

where $NLS{T}_i$ and $LS{T}_i$ denote the normalized LST and the original LST value for pixel $i$, respectively; ${LST}_{max}$ and ${LST}_{min}$ represent the maximum and minimum LST values of the winsorized LST composite in each city.

2.2.3. Google Earth Imagery

High-resolution Google Earth imagery was employed as an independent visual reference, due to its sub-meter spatial resolution and ready accessibility. These features have facilitated its extensive adoption in urban remote sensing studies [60]. For each study site, the highest-quality scenes in 2020 were selected based on minimal cloud obscuration and comprehensive coverage of the urban administrative extent. These datasets subsequently served as the primary reference for manual interpretation, the labeling of validation samples, and the precise delineation of urban boundaries.

2.2.4. Comparison Products

Two benchmark datasets were employed to assess the accuracy and robustness of the extracted urban boundaries. The first is He’s Dataset of Urban Built-up Area in China (1992–2020) V1.0 [61], which provides nationwide urban extent information at a spatial resolution of 1 km. This dataset was generated using long-term NTL observations combined with auxiliary statistical and spatial constraints to ensure temporal consistency. With a reported overall accuracy of 92.62% and a Kappa coefficient of 0.60, it is widely used for analyzing large-scale urban expansion patterns and long-term urbanization trends. The second benchmark dataset is Sun’s Dataset of Built-up Areas of Chinese Cities in 2020 [62], which delineates urban extents at a spatial resolution of 10 m based on multi-source Sentinel imagery. Its production integrates Sentinel-1 SAR data, Sentinel-2 optical imagery for vegetation index extraction, and global land cover products, followed by expert interpretation and manual refinement using high-resolution Google Earth imagery. This dataset achieves high accuracy, with an average overall accuracy of 95.57% and a Kappa coefficient of 0.91, and provides detailed and reliable urban boundary information for Chinese cities in 2020. However, since Sun’s dataset focuses exclusively on core urban areas, certain peripheral districts (e.g., Naxi District in LZ and Enyang District in BZ) are omitted from its coverage. Comparative analyses were thus limited to the spatially overlapping regions to ensure consistency and validity.

2.3. Methods

The methodological workflow implemented in this study comprised three main steps: data preprocessing, urban extraction, and accuracy assessment (Figure 2). First, the VNL data were harmonized with Landsat-derived spectral indicators and LST, all of which were preprocessed on the GEE platform, as detailed in preceding sections. These multi-source variables were integrated to construct the VWBTNUI. Subsequently, the image was classified using an unsupervised K-means algorithm to separate impervious from pervious land. The initial classification was refined through polygon aggregation and removal of fragmented patches to generate a morphologically coherent urban extent. Finally, the extracted results were evaluated using visual inspection and quantitative approaches. This included spatial pattern analysis, transect profile comparisons, and pixel-level accuracy metrics compared against raw VNL and four established NTL-derived indices. Supplemental validation was conducted via area-based assessments and cross-comparison with two urban products to determine the reliability and spatial consistency of the VWBTNUI-derived results.

2.3.1. Construction of VWBTNUI

Integrating a comprehensive and diverse feature set helps mitigate the intrinsic limitations of individual datasets and enhances classification accuracy [62]. Building on this premise, the VWBTNUI combines NTL data with NDVI, MNDWI, NDBI, and LST to capture the dominant biophysical contrasts between built-up and non-built-up environments. Each indicator represents a distinct aspect of land-surface properties. NDVI is used to suppress vegetated surfaces, assigning lower values to densely vegetated pixels and thereby facilitating the identification of sparsely vegetated impervious areas [63]. MNDWI mitigates interference from water bodies, which frequently generate spectral confusion in river-dominated landscapes [64]. NDBI enhances the representation of impervious materials, sharpening the contrast between built-up structures and natural land covers [34]. LST provides thermal differentiation by capturing the elevated heat emission of compact residential and industrial zones [65].

To integrate the multiple components, we adopted a multiplicative formulation, in which each variable is transformed into a non-negative or positive term to ensure numerical stability. Vegetation suppression is expressed as (1–NDVI), converting dense vegetation into low values, thereby highlighting sparsely vegetated areas that are more typical of urban land. Water suppression is implemented as (1 − MNDWI); this transformation guarantees non-negative values while effectively reducing the influence of water bodies. Conversely, built-up enhancement is represented by (1 + NDBI), which shifts the index into a non-negative domain and amplifies impervious surface signatures. For the thermal component, LST values are first winsorized and normalized to the [0, 1] range (NLST) as described in Section 2.2.2, and then incorporated as (1 + NLST). These transformations maintain all multiplicative terms non-negative and numerically stable. The resulting value ranges of the transformed components are summarized in

Table 3. Value ranges of components.

Term	Original Range	Transformation	Resulting Range
NDVI	[−1, 1]	1 − NDVI	[0, 2]
MNDWI	[−1, 1]	1 − MNDWI	[0, 2]
NDBI	[−1, 1]	1 + NDBI	[0, 2]
NLST	[0, 1]	1 + NLST	[1, 2]

.

Compared with additive formulations, the multiplicative approach restricts the response space, suppresses noise propagation, and promotes greater internal consistency in delineating built-up morphology [36]. The final formulation is presented in Equation (5), with each term contributing to a consolidated representation of the physical conditions inherent to urban land.

$$VWBTNUI=(1-NDVI)\times (1-MNDWI)\times (1+NDBI)\times (1+NLST)\times NTL$$

(5)

where $NTL$ is the digital radiance value of VNL.

2.3.2. Urban Extent Extraction

Urban extent delineation was executed using a two-step procedure consisting of unsupervised clustering followed by spatial refinement. While clustering provides the initial classification, post-processing is essential to eliminate artifacts and achieve a realistic representation of urban form.

(1) K-means clustering. K-means clustering was selected for the initial classification as it does not require labeled training samples or pre-defined thresholds [66], which are often unavailable for small and medium-sized cities lacking authoritative impervious surface records [67]. The algorithm partitions pixels into clusters based on feature similarity, iteratively updating cluster centroids and reassigning pixels until convergence. Formally, K-means minimizes the within-cluster sum of squares, expressed as:

$$J={\displaystyle \sum _{k=1}^K{\displaystyle \sum _{x_i\in C_k}{‖x_i-{\mu }_k‖}^2}}$$

(6)

where K is the number of clusters; $x_i$ represents a pixel vector, $C_k$ denotes the set of pixels assigned to cluster k, and ${\mu }_k$ is the centroid of cluster k. In this study, K was set to two, yielding a binary classification of urban (1) and non-urban (0) pixels. To ensure consistent and reproducible clustering results, K-means algorithm was implemented using the algorithm from the scikit-learn library (Python 3.10). Key parameters were explicitly specified (Table A1) to reduce sensitivity to random initialization. Given the binary clustering objective of the VWBTNUI values, this configuration yielded stable and consistent urban–non-urban separation across all study areas.

(2) Post-classification refinement. Although K-means provides efficient and objective classification, the initial outputs often contain isolated pixels and irregular boundaries that deviate from actual urban morphology. The classification results were therefore converted into polygon features and refined in ArcGIS 10.8 using the Aggregate Polygons and Eliminate Polygon Part tools to merge adjacent patches and fill internal holes, followed by removal of isolated patches (key parameters are listed in Table A1). All results were visually inspected, and minor manual refinements were applied solely to correct local artifacts and ensure the structural integrity of urban features. This sequential refinement strategy produced spatially coherent polygons while maintaining topological consistency.

2.3.3. Validation and Accuracy Assessment

To contextualize the performance of the VWBTNUI, four established NTL-derived composite indices were selected for comparative analysis (

Table 4. Comparison indices.

Index	Formula	Reference
BANUI	$BANUI=(1+NDBI)\times NTL$	[68]
VANUI	$VANUI=(1-NDVI)\times NTL$	[30]
VTLI	$VTLI=(1-NDVI)\times LST\times NTL$	[36]
VWANUI	$VWANUI=(1-NDVI)\times NDW{I}_b\times NTL$	[64]

Note: (a) The binary image of $NDW{I}_b$ is defined such that pixels with NDWI value less than −0.1 are assigned a value of 1, while all other pixels are assigned a value of 0. (b) No logarithmic transformation was applied to NTL values during the construction of the VWANUI to ensure direct comparability among all indices evaluated.

): the vegetation-adjusted NTL urban index (VANUI), the building-adjusted NTL urban index (BANUI), VTLI, the vegetation-water-adjusted NTL urban index (VWANUI).

The choice was based on three main considerations: (1) the inclusion of diverse correction strategies, (2) computational simplicity and interpretability, and (3) strong representativeness of key influencing factors. Socioeconomic-related indices (e.g., those relying on POIs, road networks, or geolocation data) were excluded due to their high data acquisition requirements and significant processing complexity. Similarly, indices such as HSI and NUACI were omitted owing to their reliance on complex parameters and limited interpretability [30,67]. Each chosen comparative index addresses distinct challenges in urban mapping: BANUI incorporates building information to handle morphological heterogeneity; VANUI integrates NDVI to mitigate saturation in brightly lit cores; VTLI combines vegetation and LST to refine extent detection; and VWANUI jointly suppresses vegetation and water signals to reduce blooming effects. Their inclusion provides a robust reference framework for evaluating the multidimensional improvements introduced by VWBTNUI.

Validation was conducted through both qualitative and quantitative assessments. Qualitatively, the spatial distribution of VWBTNUI was visually inspected to evaluate its ability to suppress blooming effects and to delineate fine-scale urban structures [45]. Furthermore, transect profiles were extracted across representative land-cover gradients to compare response patterns of the different indices. This analysis provides insights into the sensitivity under heterogeneous surface conditions [51,63].

For quantitative assessment, a rectangular validation zone was delineated for each city, encompassing the union of all classified urban extents. Within each zone, 250 stratified random points were generated, with an approximate 1:2 ratio of urban to non-urban samples, ensuring statistical representativeness [69]. Pixel values from each index were extracted using the Extract Multi Values to Points tool in ArcGIS 10.8. Ground truth labels were derived from Google Earth imagery via manual interpretation, with binary assignments of 1 (urban) and 0 (non-urban). Classification performance was evaluated using confusion matrices, from which the OA, F1 score, and the Kappa coefficient were calculated [64,70]. OA measures the proportion of correctly classified pixels, while the F1 score, defined as the harmonic mean of user’s accuracy (UA) and producer’s accuracy (PA), effectively balances omission and commission errors. The Kappa coefficient further accounts for agreement occurring beyond random chance. The combined use of these metrics provides a robust assessment of classification reliability. Their formulations are expressed as follows:

$$OA={\displaystyle \frac{{\displaystyle \sum _{i=1}^k{n}_{ii}}}{N}}$$

(7)

where $n_{ii}$ represents the number of correctly classified samples in class i, k is the total number of classes, and N is the total number of samples.

$$F{S}_i=2\times {\displaystyle \frac{U{A}_i\times P{A}_i}{U{A}_i+P{A}_i}},U{A}_i={\displaystyle \frac{n_{ii}}{n_{+i}}},P{A}_i={\displaystyle \frac{n_{ii}}{n_{i+}}}$$

(8)

where $F{S}_i$, $U{A}_i$, $P{A}_i$ denote the F1 score, UA, and PA for class $i$, respectively; $n_{+i}$ and $n_{i+}$ represent the column and row totals for class $i$.

$$KC={\displaystyle \frac{P_o-P_e}{1-P_e}},P_o=OA,P_e={\displaystyle \sum _{i=1}^k(\frac{n_{i+}}{N}}\times {\displaystyle \frac{n_{+i}}{N}})$$

(9)

where $KC$ is the Kappa coefficient, $P_o$ is the observed agreement and $P_e$ is the expected agreement by chance.

Beyond pixel-level assessment, area-based validation was conducted to evaluate the accuracy of urban extent delineation [71]. For each study site, reference urban boundaries were manually digitized from Google Earth imagery and independently verified by two additional specialists to minimize subjectivity. To quantify the discrepancy between the estimated and actual urban extents, the relative error (RE) was calculated as Equation (10). This metric captures systematic biases in area estimation and allows equitable comparisons across cities of different sizes.

$$RE={\displaystyle \frac{\left|S-S_{ref}\right|}{S_{ref}}}\times \mathrm{100\%}$$

(10)

where S denotes the extracted urban area and $S_{ref}$ represents the reference area.

3. Results

3.1. Spatial Distribution Characteristics of VNL and NTL-Derived Indices

Figure 3 presents the spatial distributions of the raw VNL image alongside five NTL-derived indices across all study cities. A primary observation is that VWBTNUI demonstrated the most compact and continuous high-value clusters within urban cores, while maintaining low values in non-urban regions. In contrast, the VNL data showed significant blooming effects and blurred boundaries, leading to a systematic overestimation of urban extents. Similarly, BANUI provided minimal improvement, with a spatial distribution largely consistent with VNL. While VANUI and VTLI succeeded in reducing core saturation, both indices displayed an undesirable response along river-adjacent zones, most notably in ZY. Furthermore, although VWANUI effectively delineated water bodies, its sensitive binary formulation introduced voids within dense built-up fabrics, potentially indicating localized misclassifications. Conversely, the VWBTNUI reduced the road-related blooming effects (e.g., an area reduction of approximately 5 km² along the road connecting the urban core to Enyang Airport in BZ) and captured richer intra-urban detail. Major rivers and channels were consistently identifiable, and cross-river built-up corridors remained intact (e.g., in ZY and LZ).

3.2. Transect Analysis

While spatial maps provide an overview of distribution patterns, transect analysis offers a finer-scale perspective on index sensitivity across heterogeneous land-cover gradients. Latitudinal transects were drawn across each city (Figure 4), intersecting diverse surfaces, including urban fringes, dense residential districts, urban parks, river channels, and suburban industrial zones.

Across the three sites, VWBTNUI exhibited a consistent and well-structured pattern along the extracted profiles. The index consistently maintained low values over rural and water surfaces while showing a sharp increase at the urban-rural interface and reaching high-values within core urban districts. In contrast, the raw VNL profiles were characterized by excessive smoothness and artificial peaks near major roads and adjacent bare soil patches (Figure 4a), a direct consequence of the NTL blooming effect. BANUI closely mirrored the VNL curves, producing uniformly high responses that obscured distinctions among bare soil, sparsely vegetated areas, and built-up zones due to insufficient spectral contrast (Figure 4b). Notably, VANUI and VTLI still showed elevated responses along river segments in ZY and LZ (Figure 4a,c). By incorporating water-suppression mechanisms, VWANUI and VWBTNUI effectively improved water discrimination. Although both indices followed a similar trajectory, VWBTNUI yielded noticeably lower responses over intra-urban bare land, resulting in smoother and more continuous transect profiles within compact urban zones.

3.3. Urban Extent Extraction Results

3.3.1. Spatial Patterns of Extracted Impervious Areas

Figure 5 illustrates the spatial distribution of impervious surfaces extracted using K-means clustering across the six investigated layers. Reference urban boundaries, manually delineated from Google Earth imagery, are shown for comparison. Representative subareas highlighted in Figure 5 (green circles) are enlarged in Figure 6 for closer inspection of classification patterns.

Among the evaluated approaches, VWBTNUI produced the most spatially coherent and accurate representation of urban extents. It maintained a clear distinction between urban and non-urban land, suppressed road-associated lighting (e.g., ZY, Figure 5), and delineated river channels and lakes clearly (e.g., LZ, Figure 6). In addition, VWBTNUI preserved fine-scale urban building configurations that were largely obscured in other indices (e.g., ZY, Figure 6). By contrast, raw NTL and BANUI outputs exhibited coarse morphologies and overestimated urban extents, producing widespread bright patches along illuminated traffic corridors, such as the airport access roads in BZ (blue circle, Figure 5), resulting in enlarged and spatially diffuse urban footprints. The performance of intermediate indices, such as VANUI and VTLI, showed improved intra-urban sensitivity, particularly in vegetated residential areas, yet exhibited significant degradation near river corridors. Notably, VWANUI exhibited localized fragmentation in dense urban cores (e.g., BZ in Figure 6), suggesting potential misclassification effects related to building shadows.

While a small fraction of peripheral urban zones remained undetected in the VWBTNUI results—largely due to the “dark area” problem in regions with minimal nighttime illumination (e.g., ZY, Figure 5)—this limitation is shared by all NTL-based approaches. Despite these localized omissions, the overall spatial pattern of VWBTNUI remained consistent with the reference boundaries and preserved the most continuous built-up structures across diverse inland cities.

3.3.2. Accuracy Assessment of Urban Area Extraction

Quantitative assessment through accuracy metrics (

Table 5. Classification accuracy of urban area extraction across ZY, BZ, and LZ.

Index	OA (%)			F1 Score			Kappa Coefficient
	ZY	BZ	LZ	ZY	BZ	LZ	ZY	BZ	LZ
VNL	68.8%	75.2%	73.2%	0.64	0.69	0.67	0.49	0.58	0.54
BANUI	79.2%	87.2%	85.2%	0.71	0.80	0.80	0.55	0.71	0.68
VANUI	83.6%	88.0%	87.6%	0.74	0.81	0.80	0.63	0.72	0.71
VTLI	83.6%	88.0%	88.4%	0.76	0.81	0.81	0.64	0.72	0.73
VWANUI	86.8%	88.0%	85.2%	0.78	0.79	0.77	0.68	0.71	0.68
VWBTNUI	88.0%	90.8%	89.6%	0.80	0.85	0.84	0.72	0.78	0.75

) confirmed VWBTNUI as the most robust method among all indices. It achieved the highest mean OA (89.5%), mean F1 score (0.83), and mean Kappa coefficient (0.75). In contrast, VNL exhibited the weakest performance, yielding OA values ranging between 68.8% and 75.2% and Kappa coefficients below 0.60. These results reinforce the limitation of using radiance alone for precise urban surface delineation.

Incremental improvements were observed through the integration of spectral and thermal components. BANUI and VANUI provided enhanced reliability, with OA values exceeding 80% in most scenarios. Although VTLI and VWANUI, both extended from the VANUI framework, demonstrated further gains, their performance varied notably across cities. For example, VTLI manifested pronounced accuracy improvements in LZ (OA = 88.4%, Kappa = 0.73). Similarly, VWANUI optimized classification precision exclusively in ZY (Kappa = 0.68), failing to maintain consistent performance elsewhere. These fluctuations highlight that low-dimensional feature integration may be inadequate to fully capture the structural complexity of diverse urban landscapes.

Figure 7 illustrates distinct differences in performance among the three cities. BZ achieved the highest and most stable precision across most evaluation metrics, with its index polygons consistently showing broader spatial coverage. In contrast, ZY manifested the most significant decline in precision, particularly in the Kappa coefficient (Figure 7c), where values were markedly lower than those of the other cities. The performance in LZ remained moderate and balanced, positioned between the high stability observed in BZ and the classification challenges encountered in ZY. Despite these contextual variations, VWBTNUI consistently delivered the highest accuracy and spatial consistency, providing the most reliable depiction of urban areas among all tested methods.

3.3.3. Urban Area Estimation and Error Assessment

Beyond pixel-level accuracy, area-based evaluation provides insights into the aggregate reliability of urban extent delineation. To quantify this, the extracted impervious surfaces were converted into vector format for estimating total urban area. The resulting areas were compared with manually delineated reference boundaries using RE as the evaluation metric (

Table 6. Urban area estimates and REs from different indices.

Index	Urban Area (km2)
	ZY (36.46)	RE	BZ (28.51)	RE	LZ (87.61)	RE
VNL	50.24	37.79%	66.34	132.69%	135.32	54.46%
BANUI	49.64	36.15%	61.02	114.03%	128.59	46.78%
VANUI	41.92	14.98%	35.49	24.48%	103.89	18.58%
VTLI	42.18	15.69%	38.13	33.74%	97.89	11.73%
VWANUI	38.37	5.24%	32.08	12.52%	91.12	4.01%
VWBTNUI	39.00	6.97%	31.13	9.19%	88.87	1.44%

Note: The reference urban area for each city is shown in parentheses.

).

Among all approaches, the VWBTNUI achieved the most accurate and stable urban area estimates, with RE values consistently below 10% across three study sites. In comparison, areas extracted from raw VNL showed the largest overestimations, with RE exceeding 37% in every case. BANUI partially reduced this overestimation; however, substantial errors persisted in cities affected by severe illumination spillover (e.g., BZ). More consistent reductions were observed with VANUI, which lowered RE below 25% across the three cities. Additional improvements were observed with multi-dimensional indices: VTLI reduced RE to 11–34%, while VWANUI performed more stably, keeping RE below 13% across all cities. Nevertheless, their accuracies remained lower than that of VWBTNUI. For example, in LZ, the urban area estimated by VWBTNUI (88.87 km²) closely matched the reference value (87.61 km²), highlighting its consistently low deviation and robustness against blooming effects and city-specific illumination patterns.

3.4. Comparison with Existing Urban Products

Figure 8 and

Table 7. Urban area estimates and REs from different products.

Product	Urban Area (km2)
	ZY (36.46)	RE (%)	BZ (14.8)	RE (%)	LZ (78.65)	RE (%)
He’s	45.00	23.42	23.00	55.41	93.00	18.25
Sun’s	38.25	4.91	14.74	0.41	60.68	22.85
VWBTNUI	39.00	6.97	16.01	8.18	82.94	5.45

Note: The reference urban area for each city is shown in parentheses, representing the manually delineated core urban areas.

compare urban extents derived from VWBTNUI with two widely used products: He’s dataset and Sun’s Sentinel-based product. Among the evaluated datasets, VWBTNUI showed higher agreement with reference boundaries, preserving compact morphology, accurately delineating river edges, and maintaining built-up clusters that appeared fragmented or omitted in the external datasets. Owing to its coarse spatial resolution, He’s dataset tended to overestimate urban extent and omit several compact built-up cores, including those in central BZ (white circle, Figure 8). By contrast, Sun’s product, benefiting from high-resolution Sentinel imagery, produced sharper boundaries; however, it exhibited misclassification errors, misidentifying peripheral rural settlements as urban in ZY and BZ (yellow circles). Furthermore, it failed to capture isolated industrial parcels in LZ (white circle). In comparison, VWBTNUI achieved better consistency, particularly in delineating river edges (blue circle in BZ), underscoring its robustness relative to existing products.

The quantitative comparison presented in Table 7 further corroborates these findings. VWBTNUI maintained relatively low and balanced RE values across all three cities (6.97% in ZY, 8.18% in BZ, and 5.45% in LZ), demonstrating strong cross-city stability. He’s dataset consistently overestimated urban extent, with RE exceeding 18% in all cases, highlighting its limitations in urban extent mapping. Although VWBTNUI did not attain the lowest RE in each individual city, the deviations remained small and consistent. For instance, in ZY, the VWBTNUI-derived estimate closely matched both the reference boundary and Sun’s dataset. In BZ, deviations were limited to within 1.5 km². In LZ, VWBTNUI outperformed Sun’s dataset, which underestimated the urban area by 22.85%, likely due to the exclusion of peripheral industrial zones resulting from its distance-based definition [62]. Overall, considering both spatial morphology and total urban area, VWBTNUI offers the most reliable and balanced representation among the evaluated products.

4. Discussion

4.1. Interpreting the Performance of Existing Indices and VWBTNUI

Both qualitative and quantitative evaluations indicate that VWBTNUI provides the most reliable delineation of urban extent across the three inland cities. This advantage is particularly pronounced when the extraction results are considered within the distinctive developmental context of inland urban environments. Unlike the relatively contiguous expansion observed in coastal regions, inland cities, often constrained by rugged topography, tend to develop through fragmented parcels. This “roads-first, construction-later” development trajectory frequently produces extensive illuminated corridors connecting before the built-up surfaces are fully consolidated. Consequently, radiance-only indices are prone to systematic overestimation caused by light spillover [72,73,74]. A representative case was observed in BZ, where illumination along the corridor connecting the urban core to Enyang Airport resulted in an overestimation of nearly 5 km², corresponding to a localized extraction error as high as 132.69%.

Pronounced spectral fragmentation within inland heterogeneous mosaics weakens the effectiveness of reflectance-based indices. In this study, BANUI exhibits blurred spatial patterns comparable to those of VNL, reflecting its limited discriminative capacity under mixed land-cover conditions. Statistical evidence supports this observation, as NDBI values across the study sites exhibit consistently low mean values and limited variability, as indicated by the mean and standard deviation (SD) (ZY: mean = 0.019, SD = 0.072; BZ: mean = 0.013, SD = 0.067; LZ: mean = 0.009, SD = 0.079). Such limited variability reduces separability among built-up surfaces, cropland, bare soil, and construction sites, thereby explaining why NDBI-based methods, although effective in large metropolitan areas [75], perform poorly in fragmented inland settings where SWIR–NIR contrast is substantially reduced.

Environmental interference from water bodies and urban shadows remains a persistent source of misclassification. Vegetation-adjusted indices improve urban-rural contrast. However, their sensitivity to low-NDVI features often manifests as elevated responses along river segments. This effect arises from the formulation (1 − NDVI), which mathematically amplifies signals over water surfaces where NDVI values are typically minimal. Consequently, both VANUI and VTLI manifested elevated responses along river corridors. While integration of water constraints in VWANUI and VWBTNUI alleviates this inflation, conventional NDWI remains vulnerable to shadow-driven confusion in compact inland districts. This limitation is evident in the localized fragmentation within the dense urban cores of BZ observed in the VWANUI results, where narrow streets and variable building heights generate extensive shaded areas, which spectrally resemble water surfaces [76]. Furthermore, the contribution of thermal information was found to be conditionally beneficial rather than universally effective. Although LST substantially enhanced urban separation in LZ (SD = 2.66), its discriminative effectiveness was reduced in ZY and BZ, where thermal variability is weaker, with LST SDs of 2.37 and 1.89, respectively. This finding aligns with previous studies suggesting that LST-based differentiation becomes effective only when temperature gradients are sufficiently pronounced [65].

The limitations observed in those indices demonstrate that each individual indicator captures only a partial and context-dependent subset of urban surface characteristics, and their effects are neither independent nor uniformly beneficial [25,36]. In contrast, VWBTNUI outperforms competing approaches because its comprehensive structure enforces concurrence across complementary dimensions: vegetation suppression, water masking, building-specific reflectance, and thermal distinctiveness. This multi-source agreement mechanism functions through a series of targeted cross-checks: (1) illumination-driven spillover is suppressed where NDBI, NDVI, or LST contradict urban signatures; (2) water-related and shadow issues are corrected by the MNDWI, which leverages the strong attenuation of the SWIR band to enhance separability; and (3) low-radiance industrial areas are recovered when reflectance or thermal cues indicate impervious surfaces. Notably, at industrial fringes with low radiance, the combined contribution enables VWBTNUI to detect low-radiance built-up areas that other methods fail to capture, often due to elevated surface temperatures or distinctive roofing reflectance patterns [77] as illustrated by ZY (Figure 6).

To verify the technical robustness of this integration strategy, we conducted a comprehensive input-level sensitivity analysis by introducing ±5% random perturbations to each input band (NDVI, MNDWI, NDBI, NLST, and VNL) across 50 Monte Carlo simulations. As reported in

Table A2. Sensitivity analysis of extracted urban area under random perturbations.

City	Variable	Mean Urban Area (km2)	SD Area (km2)	CV (%)
ZY	NDVI	38.98	0.16	0.41
	MNDWI	38.83	0.10	0.26
	NDBI	38.84	0.08	0.21
	NLST	38.94	0.13	0.33
	VNL	38.86	0.15	0.39
BZ	NDVI	31.26	0.14	0.45
	MNDWI	31.16	0.07	0.22
	NDBI	31.22	0.11	0.35
	NLST	31.25	0.13	0.42
	VNL	31.22	0.13	0.42
LZ	NDVI	88.81	0.19	0.21
	MNDWI	88.75	0.18	0.20
	NDBI	88.71	0.16	0.18
	NLST	88.75	0.16	0.18
	VNL	88.67	0.18	0.20

, the resulting coefficients of variation (CV) for extracted urban areas remained below 0.5% across all cities and input dimensions, indicating that VWBTNUI is highly stable with respect to small input fluctuations. Furthermore, acknowledging the potential impact of the resolution gap between sensors, we assessed the sensitivity of VWBTNUI to potential spatial misregistration of the VIIRS data. The raster was systematically displaced by one-half pixel (±250 m) in the four cardinal directions at its native 500 m resolution prior to resampling and index computation. The resulting urban area estimates showed only minor variation, with absolute relative differences below 2% across all shift scenarios and study areas (

Table A3. Sensitivity analysis of extracted urban area to spatial misregistration.

City	East (%)	West (%)	North (%)	South (%)
ZY	−0.20	0.20	0.41	−1.00
BZ	−0.40	−0.44	−0.51	−1.11
LZ	−1.55	−0.61	−0.31	−1.18

Note: Relative differences in extracted urban area (%) under ±250 m shifts of VIIRS NTL in cardinal directions. Negative values indicate a decrease compared to the baseline (Original).

), confirming that the index is resilient to sub-pixel geo-location uncertainties. Beyond input uncertainty and spatial misregistration, the robustness of VWBTNUI with respect to validation sample selection was evaluated using a bootstrap resampling procedure (500 iterations per city). As summarized in

Table A4. Bootstrap-based sensitivity analysis of classification accuracy metrics.

City	OA (Mean ± SD %)	F1 Score (Mean ± SD %)	Kappa Coefficient (Mean ± SD %)
ZY	0.878 ± 2.04%	0.800 ± 3.31%	0.723 ± 4.59%
BZ	0.904 ± 1.40%	0.854 ± 4.10%	0.782 ± 4.78%
LZ	0.898 ± 1.97%	0.842 ± 3.33%	0.752 ± 4.57%

, the small SDs (all below 5%) for OA, F1 score, and the Kappa coefficient confirm that the binary classification performance is statistically stable and not driven by particular configuration of validation samples.

While the index demonstrates high statistical stability, its operational efficacy is ultimately modulated by the spatial configuration of the urban environment. Inter-city performance disparities clearly indicate that urban morphology is a critical determinant of extraction fidelity. In ZY, the highly fragmented urban fabric—characterized by the intermixing of cropland, bare soil, and unbuilt plots—poses the greatest challenge for accurate boundary delineation. In contrast, BZ exhibits topographically constrained and spatially concentrated expansion, which reduces surface heterogeneity and facilitates higher extraction accuracy. LZ represents an intermediate case, where hydrological complexity combined with the spatial decoupling of peripheral industrial zones from the residential core introduces additional spectral ambiguities, limiting overall performance.

Despite pronounced variability in urban environmental conditions, VWBTNUI consistently maintains superior delineation performance across all cities. By jointly integrating complementary constraints on vegetation, water, building and thermal characteristics with NTL data, this index demonstrates that multi-source concurrence is essential for NTL-based urban mapping [75,78]. This is particularly relevant in inland environments characterized by fragmented morphologies and diverse surface compositions.

4.2. Practical Implications for Inland, Data-Scarce Regions and Sustainable Development

Beyond methodological performance, VWBTNUI offers substantial practical value for urban governance and sustainable development, particularly in small and medium-sized cities where accurate spatial information is often lacking. Reliable urban delineation generated by VWBTNUI provides a robust empirical basis for delineating urban growth boundaries, enforcing zoning regulations, and assessing urban compactness. These governance functions are closely aligned with the efficiency and containment objectives embedded in SDG 11.3, which require consistent measurements of land use dynamics. By accurately identifying built-up footprints within heterogeneous transitional zones, VWBTNUI reduces the risk of planning misjudgments that arise when coarse-resolution datasets fail to capture rapid urban expansion.

In addition, the proposed framework relies exclusively on open-access global datasets, including NTL and Landsat imagery, and key data preprocessing and computation were performed in the cloud-based GEE platform, making implementations financially and logistically viable for planning agencies with limited technical capacity [63]. Benchmarking tests indicate that the end-to-end processing for a city-scale area (e.g., Yanjiang District of ZY, ~1600 km²) requires less than 4 min, while a province-scale analysis (Sichuan Province, ~486,000 km²) can be completed in approximately 41 min. The unsupervised workflow further minimizes the need for training samples, manual annotation, and computationally intensive deep-learning pipelines [5]. This low-barrier implementation is particularly valuable in regions where institutions often operate under resource constraints.

Ultimately, the stable performance of VWBTNUI across heterogeneous inland environments provides a consistent and comparable basis for cross-city urban analysis and regional modeling. By offering a reproducible and transparent mapping framework, VWBTNUI supports evidence-based policymaking and enhances accountability in urban planning processes. Collectively, these attributes contribute to advancing progress toward SDG 11 in underrepresented inland regions, effectively translating methodological rigor into practical governance and sustainability outcomes.

4.3. Limitations and Future Research

Despite the promising performance of VWBTNUI, several limitations should be considered. First, the current analysis is based on a single-year snapshot (2020), which restricts direct evaluation of the index’s temporal stability and transferability. Second, the focus on three cities within a single province limits the representation of broader geomorphic and socio-economic gradients, which may constrain the generalizability of the findings to markedly different urban environments. Third, in low-brightness areas, the lack of anthropogenic lighting and the VIIRS detection threshold (0.3 nW cm⁻² sr⁻¹) limit the index’s ability to capture emerging or low-density urban areas. Although spectral and thermal indicators provide complementary information, underestimation may still occur in sparsely developed zones. Finally, the multiplicative fusion strategy may be less responsive under extreme conditions, potentially allowing uncertainties from individual input layers to propagate.

To address these limitations, subsequent work will extend the proposed framework to multi-year time series to systematically assess the temporal stability and transferability of VWBTNUI. Expanding the analysis to a more diverse set of cities will allow a rigorous evaluation of the method’s applicability beyond regionally constrained settings. Additionally, higher-resolution NTL observations from emerging platforms such as Luojia-01 and SDGSAT-1 [79,80] will be explored to improve the detection of small or dimly lit urban features unresolved by VIIRS. Methodologically, continued refinements of the fusion strategy will incorporate adaptive weighting and data-driven thresholding to improve the balance between multi-dimensional signal contributions and noise suppression.

5. Conclusions

This study introduced VWBTNUI, a multi-dimensional NTL-based index designed to address blooming and illumination heterogeneity in urban extent mapping. By integrating vegetation, water, building, and thermal information with NTL data, VWBTNUI provides a physically consistent representation of built-up surfaces and enhances classification robustness in heterogeneous inland environments.

Empirical evaluation across three inland cities in Sichuan Province demonstrated that VWBTNUI consistently outperformed VNL and existing composite indices, achieving OA values above 0.88, F1 scores over 0.80, Kappa coefficients exceeding 0.72, and RE below 10%. While BANUI, VANUI, VTLI, and VWANUI address specific distortions, they remain sensitive to bare soil, vegetation variability, weak thermal contrast, or water-related confusion, and their performance is further influenced by urban morphological complexity. In contrast, VWBTNUI’s multi-constraint design enforces concurrence among complementary urban signatures, effectively suppressing these anomalies and producing coherent urban patterns with greater adaptability across fragmented landscapes. These results highlight that reliable inland urban extraction depends on coordinated multi-dimensional integration rather than isolated indicators.

By leveraging globally accessible datasets and efficient pixel-level operations, VWBTNUI offers an easily implementable framework for urban monitoring, planning, and environmental assessment. However, to address current limitations—specifically in detecting dimly lit settlements and the constraints of limited sample sizes and timeframes—future research should integrate higher-resolution data, adaptive fusion strategies, and broader spatiotemporal validation. Overall, this study demonstrates that multi-dimensional concurrence constitutes a fundamental methodological principle for accurate NTL-based urban delineation in inland environments.