Analysis of urban expansion patterns in the Yangtze River Delta based on the fusion impervious surfaces dataset

2.1 Study area

The YRD region is a pivotal intersection of the Belt and Road Initiative and the Yangtze River Economic Belt (Figure 1), playing a critical strategic role in China’s national modernization and development [43]. It is distinguished by a high level of urbanization, densely populated built-up areas, a vibrant economy, efficient transportation networks, and advanced infrastructure [44]. As the largest and most developed urban cluster in China, the YRD not only propels economic and social progress but also functions as a vital platform for international cooperation. Geographically, the YRD is characterized by a typical river-mouth alluvial plain topography, with predominantly flat terrain covering approximately 83.7% of the total area. North of the Yangtze River, the terrain is primarily plain, while south of the river, it is predominantly hilly. Notably, Zhejiang Province exhibits complex topography, with over 70% of its land classified as mountainous or hilly. In contrast, Shanghai consists entirely of the alluvial plain formed by the YRD. Climatically, the YRD experiences a subtropical monsoon climate, marked by hot and humid summers, as well as mild and rainy winters.

Figure 1

Map of land cover and urban expansion in the YRD. The YRD includes Jiangsu Province, Zhejiang Province, Anhui Province, and Shanghai Municipality, with Jiangsu Province including Nanjing, Wuxi, Xuzhou, Changzhou, Suzhou, Nantong, Lianyungang, Huai’an, Yancheng, Yangzhou, Zhenjiang, Taizhou, and Suqian; Zhejiang Province includes Hangzhou, Ningbo, Wenzhou, Jiaxing, Huzhou, Shaoxing, Jinhua, Quzhou, Zhoushan, Taizhou, and Lishui; Anhui Province includes Hefei, Huaibei, Bozhou, Susong, Bengbu, Fuyang, Huainan, Chuzhou, Lu’an, Anqing, and Huangshan.

2.2 Land cover and IS dataset

The IS datasets utilized in the study include CLCD, GLC_FCS30, and GISA. The CLCD is an annual land cover dataset derived from Landsat imagery, offering 30 m land use analysis specific to China over a three-decade time series. GLC_FCS30 represents the first global fine-scale land cover dataset at 30 m resolution generated through continuous change detection methods, encompassing 35 refined classifications spanning from 1985 to 2022. This dataset integrates a dense time series produced from Landsat imagery via continuous change detection, local adaptive update models, and spatiotemporal optimization algorithms, achieving an accuracy of approximately 80%. GISA provides global IS maps dating back to 1985, characterized by both a long temporal span and high accuracy due to its extensive set of randomly selected validation samples obtained from third-party data. Furthermore, it introduces a novel global IS mapping method that combines semi-automatic sample collection, local adaptive classification strategies, and spatiotemporal post-processing procedures to effectively extract IS.

The validation dataset adopted the multi-temporal IS validation dataset in YRD [45]. This dataset collects 3,845 validation samples based on high-resolution images on Google Earth, including 2,160 non-impervious surfaces (NIS) samples and 1,685 IS samples spanning from 1985 to 2020.

2.3 Data fusion method

The fusion problems in the D-S evidence theory are tackled within the recognition framework U, which is a non-empty set comprising a series of mutually exclusive elements A, and the power set of this recognition framework contains a total of 2^U elements. All objectives in the problem domain involve set functions m: 2^U → [0, 1] that satisfy

where m represents the Basic probability assignment (BPA) function for the recognition framework U. m(A) denotes the BPA for the proposition A, reflecting the degree of support for its occurrence. Equation (1) specifies that the BPA for the empty set is zero. Equation (2) indicates that the total confidence assigned to all propositions within the recognition framework sums to 1, despite each proposition having its distinct level of confidence.

Different sources of evidence m 1 , m 2 , … m n have distinct BPA functions. The D-S synthesis formula utilizes orthogonal sums to combine these individual BPA functions into a new function. This formula is expressed as follows:

where k is defined as the conflict coefficient. A conflict coefficient (k) nearing 1 indicates higher degree of conflict among the evidence sources, while k tending toward 0 signifies greater consistency among them. The term 1 1 − K serves as a normalization factor, ensuring that data fusion does not assign confidence to the empty set.

In the study, the BPA values for the classification of IS vs NIS were first determined by calculating the confusion matrix for each dataset. Then, equation (3) was implemented to fuse the evidence from the three datasets at a pixel level, which generated a composite belief for the classification of each pixel. Prior to fusion, a consistency analysis of the three types of land cover data was performed. Based on the spatial consistency analysis results, if the land cover types from all three datasets or two of them are consistent, the land cover type of the pixel was defined accordingly. In contrast, if each dataset suggests a different type of land cover for the same pixel, a standard fusion process was applied. When the belief assigned to the “impervious surfaces” class for a given pixel was maximal, it was designated as an impervious pixel, facilitating the final extraction of urban regions.

2.4 Urban expansion intensity index (UEII)and urban expansion differentiation index (UEDI)

UEII indicates the rate of change in urban land area over a specified period, reflecting both the speed and magnitude of increases or decreases in urban land extent. Its calculation formula is presented in equation (5).

where U i t 2 and U i t 1 are the urban land area of the city i at time t 2 and t 1 , Δ t is the period of the study. The urban area was extracted based on IS data following the methodology [24,46].

The UEDI represents the ratio of a city’s urban land expansion rate to the overall urban land expansion rate within the study area. This index quantifies the relative dynamics of urban land development in a specific city compared to the entire study region. By providing a standardized quantitative measure, UEDI enables more precise and comparable assessments of urban land expansion rates across different cities. Furthermore, it reflects the coordinated development and relative progress of urban expansion among cities within the region, which is essential for analyzing spatial distribution and dynamic evolution patterns. The calculation formula for UEDI is presented in equation (6).

where U t 2 and U t 1 are the urban land areas of the study area at times t 2 and t 1 .

2.5 Type of urban expansion

The urban expansion type is classified into three forms: enclave expansion, infilling expansion, and edge expansion [47]. Enclave expansion represents a prevalent form of urban growth during the urbanization process, characterized by the transformation of rural and other non-urban areas into urban land through development activities. In contrast, infilling expansion primarily occurs in contexts of intensive land use, where previously developed but underutilized areas within the city are redeveloped. Edge expansion refers to the outward growth of urban areas as cities expand their boundaries. The formula [48] for calculating urban expansion type T is presented as follows:

where L P represents the perimeter of a newly developed urban patch, and C is the length of the common boundary of this newly developed urban patch and the existing urban land. The urban expansion type is identified as enclave expansion when T = 0, as edge expansion when 0 < T ≤ 0.5, and as infilling expansion when T > 0.5. The steps of the study are shown in Figure 2.

Figure 2

The flowchart of the study.

3.1 Accuracy of fused IS datasets

The accuracy of the fusion dataset and the three datasets of the CLCD, GLC_FCS30, and GISA in the YRD are presented in Table 1. Among the 3,845 randomly sampled points, 2,160 were classified as NIS, and 1,685 as IS. For the CLCD dataset, 38 NIS points were incorrectly classified as impervious, resulting in a misclassification rate of 1.76%. Conversely, there were 374 instances where IS were misclassified, leading to a significantly higher misclassification rate of 22.2%. For GLC_FCS30, only 11 NIS points were incorrectly classified as impervious, with a misclassification rate of 0.51%, whereas 249 IS points were misclassified, yielding a rate of 14.78%. Although this was lower than that observed in CLCD, it remained considerably higher than the misclassification rate for NIS. By contrast, GISA demonstrated superior performance with only 9 instances of NIS being misclassified as impervious (a rate of 0.41%) and a total of 77 IS points misclassified, corresponding to a misclassification rate of 4.57%. Among the three datasets analyzed, GISA exhibited the lowest proportion of misclassified IS compared to the CLCD and GLC_FCS30 datasets.

The fusion dataset revealed that 12 NIS points had been inaccurately classified (0.56%), while 55 IS points were misclassified at a rate of 3.26%, which was notably lower than other datasets. Table 2 illustrates the overall accuracy (OA) and kappa coefficient for all datasets. Specifically, CLCD demonstrated the lowest classification efficacy, attaining an OA of 0.893 and a kappa coefficient of 0.777. In contrast, GLC_FCS30 exhibited moderate improvement with an OA of 0.932. Notably, GISA outperformed all initial datasets, achieving an OA of 0.978 and a kappa coefficient of 0.954. The fusion dataset realized the highest results, reaching an OA of 0.983 and a kappa coefficient of 0.965.

Among the 374 misclassified sample points in the CLCD dataset, 362 were identified as cropland, 6 as grassland, 5 as woodland, and 1 as waterbody. Similarly, among the 249 misclassified points in the GLC_FCS30, 236 were categorized as cropland. The complexity of land use types, the diversity of crop species, and variations in crop growth stages result in significant variability in the spectral characteristics of cropland observed in remote sensing imagery, leading to pronounced hetero spectral phenomena that complicate feature extraction. Furthermore, urbanization transitions from natural to artificial surfaces often involve complex surface compositions, further exacerbating classification uncertainty. The fusion of multiple datasets can substantially improve the accuracy of IS classification.

To investigate and elucidate the performance advantages of D-S evidence theory in data fusion, the study conducted a comparative analysis of several mainstream fusion strategies (Table 3) [49,50,51,52]. The D-S evidence theory significantly outperformed alternative fusion methods, yielding an outstanding OA of 98.26% even when relying on lower-resolution data.

3.2 Temporal evolution in urban land expansion intensity

The intensity of urban land expansion in the YRD exhibited distinct stages (Figure 3), following a trend characterized by Slow Growth, Rapid Growth, and subsequent Rapid Decline. The UEII gradually increased from 0.065 to 0.091 between 1985 and 2000, peaking at 0.132 during the period from 2000 to 2005. Subsequently, it declined to 0.041 from 2015 to 2020, indicating that the YRD urban agglomeration is progressively transitioning towards more coordinated and synergistic development.

Figure 3

The overall UEII in the YRD from 1985 to 2020.

Between 1985 and 2020, urban land areas in various cities exhibited continuous growth, though expansion intensity varied significantly across development stages. The temporal trend of the UEII for most regions in Zhejiang Province and Suzhou, Wuxi, and Changzhou in Jiangsu Province, as well as Shanghai, aligned with the overall trajectory of the YRD, peaking between 2000 and 2005. Notably, Jinhua recorded a UEII of 0.323, while Suzhou approached 0.3. In contrast, the UEII growth rate in northern Jiangsu and much of Anhui was relatively subdued, reaching its maximum value between 2005 and 2010. Xuancheng (Anhui) recorded a peak UEII of 0.522 during this period. Additionally, Yancheng, Bengbu, Chuzhou, and Fuyang attained their highest values between 2010 and 2015, while Wenzhou peaked earlier from 1995 to 2000. Overall, Jinhua demonstrated the highest expansion intensity at an impressive rate of 1.404%, followed by Shaoxing, Huzhou, and Suzhou.

The southern YRD, particularly the southeastern region, has demonstrated a higher rate of development compared to the northern area (Figure 4). Over time, the development rate in the north has begun to increase, gradually shifting the center of regional growth northward. Between 1985 and 1990, urban expansion across the YRD was relatively uniform, as evidenced by low UEII levels, with most regions reporting UEIIs below 0.1. Jinhua stood out with the highest UEII (0.171) and UEDI (2.645), primarily due to administrative adjustments. From 1990 to 2005, significant disparities in urban expansion became evident across the region, with the southeast exhibiting markedly higher UEII than those in the north. Jinhua and Shaoxing in Zhejiang Province reported UEDI exceeding 2, reflecting nearly a tenfold increase in urban area during this period. Between 2000 and 2005, the north-south disparity in urban expansion within the YRD became increasingly pronounced. In the northwestern area, only Hefei exhibited a UEDI of 0.810, while the other cities in this region reported UEDI values below 0.750. Conversely, in the southeastern area encompassing Jinhua, Shaoxing, and Suzhou, all cities recorded UEDI values exceeding 2. Additionally, most other cities had UEDI scores above 1, indicating a significant trend of urban expansion. For instance, Suzhou’s UEDI increased by nearly 1.5 times during this period. Between 2010 and 2015, the UEDI for Fuyang and Lu’an in the northwestern region of the YRD exceeded 2.0, whereas the UEDI for Shanghai and its surrounding cities was comparatively lower. Notably, Wuxi recorded a minimum value of 0.469.

Figure 4

The UEDI in the YRD from 1985 to 2020. (a) 1985–1990, (b) 1990–1995, (c) 1995–2000, (d) 2000–2005, (e) 2005–2010, (f) 2010–2015, (g) 2015–2020, and (h) 1985–2020.

3.3 Analysis of urban development patterns

From 1985 to 2020, urban land expansion in the YRD was primarily characterized by edge expansion (Figure 5), accounting for approximately 74.7% of the total expansion. Enclave expansion constituted about 19.6%, while infilling expansion was minimal at only 5.6%. During the period from 1985 to 2000, infilling expansion remained low, particularly between 1985 and 1990 when it was less than 1%. However, after 2000, edge expansion began to decline, while the proportion of infilling expansion gradually increased. By 2015–2020, infilling expansion had risen to approximately 16.33%. A significant shift was observed in Shanghai between 2000 and 2005, where edge expansion accounted for approximately 85%, while infilling expansion was only 3.5%. However, by 2015–2020, edge expansion decreased to around 51%, while infilling expansion surged to approximately 38.4%, marking a substantial increase.

Figure 5

The proportional composition of the three urban expansion types in the YRD. Each dot represents a city, with its size indicating the scale of urban expansion. (a) 1985–1990, (b) 1990–1995, (c) 1995–2000, (d) 2000–2005, (e) 2005–2010, (f) 2010–2015, and (g) 2015–2020.

To analyze the trends in urban spatial expansion types within the YRD from 1985 to 2020, Ward’s Euclidean squared distance clustering method based on the proportions of each city’s expansion types was employed. The analysis resulted in three categories: Category I represents a transition toward spatial agglomeration, characterized by a rapid decline in enclave expansion to below 10%, while edge expansion also experiences a slight decrease, remaining below 65%. Conversely, infilling type growth accelerates quickly to reach 25%, indicating a shift toward connotative development. Notable cities in this category include Shanghai, Nanjing, Hangzhou, Suzhou (Jiangsu), Wuxi, Zhenjiang, Shaoxing, and Jinhua. Category II is defined by rapid spatial growth, where the proportion of enclave expansion slightly decreases but remains elevated at above 20%. These cities continue to experience rapid outward growth with spatial diffusion still prevailing, suggesting that significant distances remain before achieving urban spatial agglomeration. Representative cities include Xuzhou, Nantong, Suzhou (Anhui), Wenzhou, Zhoushan, Wuhu, Chuzhou, Chizhou, and Xuancheng. Category III pertains to edge-expanding cities where enclave proportions decline slightly to below 15%, while edge expansions rise significantly to approximately 70%. The infilling type also shows a minor reduction to under 15%, indicating that edge expansion is the predominant mode of development. Representative cities in this category include Yancheng, Hefei, Anqing, and Taizhou; these areas continue their outward spread, reflecting an absence of intensive and compact urban development.

4.1 Interpretation of spatiotemporal dynamics and driving Forces

The spatiotemporal dynamics of urban expansion revealed in the study – characterized by a region-wide structural transition from dispersed leapfrog growth to consolidated edge-expansion and infilling – not only corroborate the established “diffusion-coalescence” theoretical framework [53] but also reflect the spatial manifestation of a convergence of powerful driving forces at work over the past three decades. This observed trajectory, including the conspicuous northward shift of the developmental center of gravity [54], is not a random phenomenon but rather the geographical expression of intricate interactions among economic, demographic, and political-institutional forces. Robust economic expansion, fueled by capital accumulation, generated significant demand for industrial, commercial, and residential land, serving as the foundational impetus [55]. Concurrently, demographic pressures, intensified by large-scale rural-to-urban migration, gave rise to a self-reinforcing feedback loop between population aggregation and territorial expansion [56].

Foremost among these forces, state policy has played a pivotal role, acting as a decisive mechanism in modulating the pace, scale, and configuration of urban expansion. The “land finance” regime, a central component of local governance, clearly accelerated rapid and often fragmented land development during the earlier stages [57]. The more recent deceleration of expansion rates and the increasing prevalence of infilling patterns also correspond to a broader shift in national macro-planning, which has pivoted from prioritizing sheer growth to championing ecological civilization and high-quality development [58]. This policy-driven evolution suggests that the YRD’s experience may serve as a precursor to the trajectories of other urbanizing regions in China [57]. Moreover, top-down regional integration strategies have directly promoted the functional decentralization of core cities, thereby contributing to the observed geographical shift in development focus [56]. Therefore, the YRD’s urban expansion is a contingent outcome shaped by the interplay between market dynamics and state regulation, with its historical evolution reflecting the dynamic interplay and changing dominance of these driving forces.

4.2 High-precision impervious surface mapping challenges and policy applications

A rigorous academic inquiry necessitates a critical appraisal of the study’s limitations, which highlight essential directions for future research. Methodologically, while the D-S theory produced high-performance results, its inherent constraints in managing high-conflict evidence and the potential subjectivity in BPA call for a systematic comparative evaluation against other advanced fusion algorithms, such as ensemble learning, modified D-S rules, or deep learning-based fusion networks [59]. Regarding data resolution, the 30 m resolution, while effective for macroscopic pattern detection, inherently limits the capacity to resolve fine-grained intra-urban textural details. Future investigations would be substantially enriched by integrating higher-resolution remote sensing data (e.g., from Sentinel or Gaofen series). Finally, the predominantly qualitative assessment of driving forces should be replaced by quantitative, spatially explicit modeling approaches, enabling a more rigorous analysis of the spatially heterogeneous impacts of diverse drivers [60].

Notwithstanding the methodological considerations, the high-precision IS area dataset generated herein offers substantial scientific value and broad policy relevance. For urban planners and local governments, the dataset can provide critical support for more accurate urban growth forecasting, evaluating the effectiveness of existing plans, and optimizing future land-use configurations and infrastructure allocation, thereby fostering a transition towards more compact and efficient urban development models. For regional development and environmental protection agencies, the expansion patterns and shifts in the center of gravity revealed herein can inform the formulation of more targeted regional coordination policies and ecological conservation strategies, such as defining urban growth boundaries and protecting critical ecological spaces. Additionally, this dataset can serve as a foundational data layer for researchers investigating the urban heat island effect, changes in biodiversity, carbon cycling, and other environmental and socioeconomic consequences of urbanization, thus facilitating deeper interdisciplinary investigations.