(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Mapping urban dynamics in Greater Sydney – A scalable multi-decadal local climate zone classification approach [1]

['Shankar Sharma', 'Climate Change Research Centre', 'Unsw Sydney', 'Sydney', 'New South Wales', 'Arc Centre Of Excellence For Climate Extremes', 'Arc Centre Of Excellence For Century Weather', 'Jason P. Evans', 'Andy Pitman', 'Negin Nazarian']

Date: 2025-09 

Abstract Urbanization has significantly altered land surface properties, leading to changes in local micro-climates and impacts on regional climate. To understand these changes, long-term, precise urban land use and land cover (LULC) data are essential. The Local Climate Zone (LCZ) classification has emerged as a high-resolution urban LULC mapping scheme, which classifies urban areas and their natural surroundings into 17 classes. However, previous efforts have primarily focused on single-year LCZ mapping. To address the lack of temporal changes, this study introduces a new approach for multi-year LCZ classification using the World Urban Database and Access Portal Tools (WUDAPT) framework. We generated the first high-resolution, multi-year LCZ maps for the Greater Sydney Region from 1990 to 2020 at five-year intervals, with each map achieving over 65% classification accuracy, surpassing the accuracy threshold set by WUDAPT’s workflow for urban land cover reliability. The LCZ map reveals significant urban expansion and densification across Sydney, particularly after 2005. The total area covered by compact and open mid and highrise LCZ classes (LCZ 1–5) has increased significantly, with growth ranging from about 56% (LCZ 1) to 493% (LCZ 5). Meanwhile, open lowrise (LCZ 6) and sparsely built (LCZ 9) areas declined by 23.84% and 15.44%, respectively. This shift toward denser urban development underscores the utility of multi-year LCZ mapping to track urban dynamics over time. Transferable to any global city, this multi-decadal LCZ mapping approach improves urban land cover accuracy at high-resolution and can enhance the precision of urban simulations. These LCZ maps support application in urban planning, climate adaptation, environmental assessments, and urban meteorological modelling, providing researchers, policymakers, and urban planners with valuable tools to address climate resilience and urban growth management effectively. Future work can integrate these LCZ datasets with climate and hydrological models to assess the long-term impacts of urban expansion on extreme weather, heatwaves, and flood risk, and apply this classification approach to other cities for comparative urban studies.

Citation: Sharma S, Evans JP, Pitman A, Nazarian N, Lipson MJ, Lopez-Bravo C (2025) Mapping urban dynamics in Greater Sydney – A scalable multi-decadal local climate zone classification approach. PLOS Clim 4(8): e0000677. https://doi.org/10.1371/journal.pclm.0000677 Editor: Ahmed Kenawy, Mansoura University, EGYPT Received: December 7, 2024; Accepted: June 27, 2025; Published: August 8, 2025 Copyright: © 2025 Sharma et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: Urban descriptive parameters are provided in the supplementary file. The LCZ classification tool can be freely accessed at https://github.com/matthiasdemuzere/multitemporal-lcz-mapping/tree/master. Landsat datasets were obtained from Google Earth Engine (https://earthengine.google.com/). The generated training polygons and Local Climate Zone maps of the Greater Sydney Region are available upon request from the lead author. Figures were created using a Python program, and the code is freely accessible at: https://github.com/ss0293/Mapping-Urban-Dynamics-in-Greater-Sydney. Funding: This work was supported by the Australian Research Council (ARC) Centre of Excellence for Climate Extremes (CE170100023 to S.S., J.P.E., A.P., N.N., C.L.-B.), and the ARC Centre of Excellence for Weather of the 21st Century (CE230100012 to S.S., J.P.E., N.N., M.J.L.). Competing interests: The authors have declared that no competing interests exist.

1. Introduction Urbanization, one type of land use land cover (LULC) change, is defined as the process of transforming natural landscapes into structures such as buildings, roads, and infrastructure [1]. Urban areas are hotspots of energy consumption and pollution emissions and can significantly impact the local environment, including weather and climate [2–4], due to their physical processes and human activities [5,6]. The global population has increasingly shifted towards urban areas, with 56% of the total global population currently residing in urban regions and this is projected to increase to 68% by 2050 [7]. This increasing population density and extensive LULC change contribute to the formation of urban heat [8,9], which ultimately intensifies the vulnerability of large populations to extreme heat events and other climate-related risks [10]. Given the intricate nature of urban landscapes and their impact, accurate urban classification is essential for understanding influences on weather and climate extremes, urban climate and urban planning applications [11,12]. In regional climate modeling, accurate land cover information is required for urban climate and weather simulations [13]. However, many widely used land cover datasets are too coarse to accurately represent the complexity of urban areas. Additionally, the previous generation land cover data generally categorizes urban areas as a single class of dense built-up regions, lacking a diverse range of urban form, fabric, and function [14]. To address this challenge, the Local Climate Zone (LCZ) classification scheme has emerged globally to provide detailed urban land cover information [15–20], including urban canopy parameters [14]. The LCZ land cover scheme has been adopted in urban climate modeling; to study the extreme rainfall [21,22], urban heat islands (UHI) effect [23–25], monitoring the daily temperature range [26] and heatwaves [27,28]. The LCZ classification scheme, originally proposed by Stewart and Oke, [9] and later adopted by the community-driven World Urban Database and Access Portal Tools (WUDAPT) initiative to generate high-resolution, class-based LULC maps for global cities [16,29]. The LCZ scheme provides a standardized framework for classifying neighborhoods within a city based on similar urban cover types, structures, materials, and human activities, thereby capturing areas with comparable UHI effects [9,16,26]. By categorizing urban areas into 17 distinct classes—comprising ten built-up urban classes and seven natural land cover classes—the LCZ scheme allows for incorporating urban climate parameters such as sky view factor and building surface fraction [30]. This framework enhances urban climate model descriptions [15,26] and serves as a valuable tool for urban morphology analyses, human health studies, and socioeconomic research [19]. While the LCZ scheme can potentially enhance urban climate simulations, its accuracy in representing urban land cover remains a significant issue, as it depends on the classification approach, methodology, and data applied [31]. To generate the LCZ maps in different cities, studies have used different methods: Geographic Information System (GIS)-based remote sensing [32–34], an Object-Based Image Analysis (OBIA) approach [35–37], and WUDAPT workflow [25,26,38,39]. The GIS-based methods typically integrate various spatial data layers to delineate different LCZs based on their physical and socio-economic characteristics [34,40]. The OBIA divides imagery into coherent objects or groups of pixels, allowing for the classification of urban forms that align well with LCZ types by considering spectral, spatial, and textural characteristics within these segments. The most widely adopted is WUDAPT’s workflow, which generates LCZ maps by using freely available satellite imagery, user-defined training polygons for each LCZ class, and a supervised classification algorithm to classify urban and natural land cover, making it accessible and adaptable for global applications [16]. Using this approach, Bechtel and Daneke, [29] achieved classification accuracies of up to 97% for Hamburg, Germany. Since then, several studies have applied similar methods to create LCZ maps for cities such as those in China [35,41], Australia [24,27,42], Hong Kong [34], the United States [43], and many European cities [26,38,44,45]. Many recent studies have adopted this approach, incorporating post-processing techniques and newer satellite datasets (e.g., Sentinel) to enhance classification [46,47], as the WUDAPT scheme primarily relies on Landsat imagery due to its long-term availability. However, these improvements often require additional data—such as building height, perviousness, sky-view factor (SVF), or canyon height-width ratio (HW)—which are not easily or freely accessible. The use of LCZ maps to improve model performance over cities is effective [25,26,37,47]; however, the majority of these studies have examined the LCZ maps at a single point (single-year scope) in time and do not capture the continuous urban expansion and densification in cities. Such static classifications are insufficient for understanding the temporal evolution of urban areas, which is necessary for assessing the cumulative impacts of urbanization on climate and the environment. The long-term urban LULC change can reveal trends, patterns, and shifts in urban areas, highlighting the necessity for a dynamic approach to urban classification [48]. Moreover, the extended temporal perspective can enhance the accuracy of urban climate models to reproduce and predict climatic conditions and urban effects reliably. There has been progress in multi-year LCZ mapping by extending classifications to historical periods. For example, Cai et al. [41] focused on three Chinese megacities and extended their classification to 2000, modifying the LCZ scheme by grouping similar functional areas (merging and removing LCZ classes) to simplify analysis for urban mapping and planning applications. More recently, Qi et al. [43] adopted a hybrid modeling approach to generate national-scale LCZ maps for the continental United States from 1986 to 2020. However, both studies relied heavily on post-processing that incorporated socio-economic data, which may introduce additional complexity and potential inconsistencies across global cities, particularly in regions where such data may be unavailable or less reliable. There is a need for scalable methods that generate long-term LCZ maps based on global satellite imagery. To address this, our study provides the first multi-year LCZ map of the Greater Sydney Region (hereafter referred to as Sydney) as a use case, covering the period from 1990 to 2020. Utilizing publicly available Landsat images and a machine learning-based classification scheme, we introduce a new approach to classify multi-decadal LCZ maps following the WUDAPT framework. Specifically, this study aims to address two research questions: (1) is it feasible to generate scalable multi-decadal LCZ maps? and (2) how effective is this method in capturing the temporal dynamics of urban growth in Sydney? The resulting LCZ maps can serve as a valuable dataset for a range of applications, including urban climate simulations and urban planning.

2. Methodology 2.1 Greater Sydney region Over the past few decades, Sydney has undergone significant urbanization, with population growth from 3.6 million in 1990 to 5.2 million in 2020 [49]. This region, located in south-eastern Australia, covers 12,367 km², stretching from the Blue Mountains in the west to the Pacific Ocean in the east, and from the Hawkesbury River in the north to the southern edge of Botany Bay. Sydney, the capital of New South Wales (NSW) and the region’s urban core, has experienced rapid economic growth, contributing to its transformation into one of Australia’s most densely populated cities. This urbanization process has driven substantial changes in land use, reshaping the morphological features and increasing the density of built-up infrastructure [50]. Sydney is characterized by diverse landscapes, including coastal zones, large national parks/natural forest areas, and mountain ranges, contributing to varied urban forms across the region. This diversity, combined with Sydney’s dynamic growth, makes it an ideal case study for examining urban transformation through time. 2.2 Multi-year LCZ classification In this study, we used a machine learning and remote sensing-based approach to produce a multi-decadal LCZ map for Sydney, covering from 1990 to 2020 at 5-year intervals. The LCZ map generation process (shown in Fig 1) involved three key steps: (1) generating training area (TA) polygons (Fig 1(step 1)), (2) LCZ classification by applying machine learning (random forest (RF) classifier) with a LCZ classification tool (Fig 1(step 2), and (3) validating the resulting LCZ maps (Fig 1(step 3)). A simplified schematic of the workflow is also provided in the Supplementary Information (S1 Fig). PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 1. Methodological framework for multi-decadal LCZ classification in Sydney. The process includes (1) manual generation of training polygons, (2) LCZ classification using Landsat images and machine learning, and (3) visual and statistical validation, where OA represents the overall accuracy or the percentage of correctly classified pixels out of the total pixels. Each colored box represents a distinct step in LCZ generation. Final TAs refer to the revised training areas used to generate the final LCZ maps. In the multi-year classification, 2020 TAs are overlaid on 2015 imagery, revised/remove and added TAs to reflect changes. These updated TAs then proceed through steps (1) to (3). The final TAs for 2015 are then used to classify 2010, and so forth for earlier years. Base map and data from OpenStreetMap and OpenStreetMap Foundation. https://doi.org/10.1371/journal.pclm.0000677.g001 2.2.1 Generate training area polygons. We defined our Region of Interest (ROI) to cover Sydney. Initially, we searched for available pre-defined Training Areas (TA) polygons within the ROI on the WUDAPT website (https://lcz-generator.rub.de/submissions). We selected some representative TA polygons from Nazarian [51], that corresponded to the land cover of Sydney in 2020. Following the guidelines provided by WUDAPT (https://www.wudapt.org/digitize-training-areas/), we created a set of TAs for 2020 using Google imagery in Google Earth Pro. We started with 2020 to leverage the benefits of Google Street View, Google Maps, local observations, and updated Landsat imagery. This process involved carefully inspecting sample areas and drawing polygons to cover representative regions for each LCZ class. Each TA was then assigned to its respective LCZ class, with further verification from Google Street View, Google Maps, and field observations. We excluded the lightweight lowrise class (LCZ 7) from classification due to its rarity in Sydney, resulting in a total of 16 land cover classes (nine urban and seven natural LCZ) within the LCZ framework, as outlined by Stewart and Oke [9] and Bechtel et al. [52]. 2.2.2 LCZ classification and quality control. After generating the complete set of TA polygons for 2020, we used a Python-based LCZ classification tool developed by Demuzere et al. [53], initially tested over Canadian cities [54] and has been developed further for Chinese megacities [41]. This classification step consisted of two stages: First all the TAs (year 2020) were uploaded to Google Earth Engine (GEE). Within GEE, latest Landsat scenes were retrieved from a 180-day window before and after the selected year, with cloud cover threshold of less than 20%. These scenes provide earth observation features that alongside the digitized TAs, served as the input data. In the final stage, a RF classifier was applied to generate the LCZ map, leveraging reflectance data from key Landsat spectral bands, including blue, green, red, near-infrared (NIR), shortwave infrared (SWIR) 1 and 2, and thermal infrared (TIRS) [45,54]. To enhance classification accuracy and better differentiate LCZ classes, a suite of derived band ratios and indices was incorporated. These include the Biophysical Composition Index (BCI), which integrates brightness, greenness, and wetness to highlight urban biophysical properties; the Normalized Difference Anthropogenic Built-up Index (NDABI) and Normalized Difference Built-up Index (NDBI), both emphasizing built-up areas critical for urban LCZs (e.g., LCZ 1–10); and the Enhanced Built-up and Bareness Index (EBBI), which leverages thermal data to refine detection of built-up and bare soil regions. Additionally, the Normalized Difference Vegetation Index (NDVI), and Normalized Difference Water Index (NDWI) were included to distinguish vegetation and water features, essential for natural LCZs (e.g., LCZ A–G). Slope data, derived from the ASTER Global Elevation Dataset (GDEM), was also added to capture topographic influences. These features were processed consistently, with cloud masking applied to ensure data quality and Gaussian filtering used to enhance spatial coherence, before being input into the RF classifier. The resulting 100 m LCZ map and accuracy statistics were generated following the methods of Demuzere et al. [53]. In this approach, 70% of the TAs were randomly selected for model training, with the remaining 30% used for accuracy assessment, and this procedure was repeated over 25 bootstrap iterations. However, all TAs were used for the final LCZ map. The accuracy metrics include overall accuracy (OA), measuring the percentage of correctly classified pixels; urban LCZ accuracy (OAu), which focuses on accuracy within urban classes; built vs. natural LCZ accuracy (OAbu), distinguishing urban from natural zones; and weighted accuracy (OAw), which applies higher penalties to errors between dissimilar classes (e.g., urban vs. natural) (Fig 2). For further details, refer to Bitchel et al. [52] and Demuzere et al. [16]. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 2. LCZ map accuracy metrics from 1990 to 2020, including overall accuracy (OA), overall accuracy for urban classes (OAu), overall accuracy for built versus natural classes (OAbu), and weighted accuracy (OAw) on the top panel across the given time periods. The bottom panel shows the distribution of training area polygons across LCZ classes for each year, with the x-axis indicating the year and the corresponding total number of training areas (TAs) for that year. The colour in the histogram (lower panel) represents the TAs from different LCZs shown in LCZ map accuracy metrics from 1990 to 2020, including overall accuracy (OA), overall accuracy for urban classes (OAu), overall accuracy for built versus natural classes (OAbu), and weighted accuracy (OAw) on the top panel across the given time periods. The bottom panel shows the distribution of training area polygons across LCZ classes for each year, with the x-axis indicating the year and the corresponding total number of training areas (TAs) for that year. The colour in the histogram (lower panel) represents the TAs from different LCZs shown in Fig 3 . Here, 1 to 10 represent urban LCZ, and A to G represent natural LCZ. https://doi.org/10.1371/journal.pclm.0000677.g002 2.2.3 Validation and decadal LCZ mapping. Following the initial generation of the 2020 LCZ map, we conducted a qualitative inspection using Google Earth Pro imagery, leveraging local insights into Sydney neighbourhoods to identify misclassified or poorly classified areas. Along with visual inspection, the map was required to meet an accuracy threshold of OA ≥ 0.6—higher than the WUDAPT standard of 0.5—as we observed that an OA below 0.6 could sometimes include misclassified pixels. If the map did not meet this threshold, we inspect the misclassified areas, added and revised TAs, and iteratively repeated the process until the OA exceeded 0.6, ensuring alignment with Google Earth imagery for the specified year. Once the final LCZ map for 2020 was generated (passing visual inspection with OA ≥ 0.6), we used 2020 TAs along with the Google timelapse feature in the Google Earth Pro platform to review the area as it appeared in 2015. We further inspected each 2020 TA polygon, and if the area’s building type or land cover remained unchanged, we retained the TA. If the area had changed, we modified the TA accordingly. If the area no longer belonged to the same class, we closely inspected it, and if we were confident enough to reclassify, we reclassified the area; otherwise, we removed the TA. Then, we followed the same process as in 2020. After generating the LCZ map, we reassessed its overall accuracy through cross-validation using Google Timelapse imagery corresponding to the same period. In cases where significant misclassifications were identified, we employed the same corrective process as was used for the year 2020. Upon achieving a classification accuracy higher than 0.6 and producing a satisfactory LCZ map (map cross-validation with Google Image), we finalized the LCZ map for 2015. This process was similarly applied to the years 2010, 2005, 2000, 1995, and 1990, with LCZ TAs from subsequent years serving as a reference for each classification (S1 Fig and Fig 1; multi-year classification process). Misclassification of pixels frequently occurred along coastal lines, where rocks or stones (LCZ E) are classified as buildings (mostly LCZ 3). To address this, we added new TA polygons (LCZ E) in these areas, which reduces misclassification significantly. Additionally, some water bodies dry up periodically and are sometimes classified as built pixels. We corrected this through post-processing, determining the maximum lake size, and overlaying as a water (LCZ G) onto all maps for consistency. This approach helps resolve misclassification errors in the LCZ map. However, it’s crucial to ensure such water bodies needs remain consistent in visual verification throughout the study period, as some water bodies are converted into built-up classes. This thorough TA inspecting and tracking process (modifying, adding, and removing), which significantly reduces the chances of misclassification in multi-decadal LCZ maps. The total number of labelled TA polygons for the years 2020, 2015, 2010, 2005, 2000, 1995, and 1990 were 642, 609, 641, 584, 550, 418, and 409, respectively (Fig 2). Notably, this study represents the first known attempt to conduct a long-term LCZ classification with independent validation over Sydney. After completing the LCZ mapping, we analysed the temporal changes in urban land cover within Sydney by calculating the area for each class during the study period. This analysis highlights the land cover changes over time and evaluates the model’s ability to capture these changes. Additionally, we calculated the transition matrix to illustrate how each class transformed into another from 1990 to 2020. The transition matrix, a quantitative tool in this study, is used to track how each LCZ pixel has changed over time. For example, if a pixel was classified as LCZ A (dense forest) in one year and LCZ 1 (compact highrise) in a subsequent year, the transition matrix captures this change, providing a detailed overview of the spatial dynamics of land cover change. We further calculated the city descriptive information following the study of Lipson et al. [31]. The calculated urban cover maps provide important inputs for urban climate models, including surface cover, morphology, canopy attributes, and thermal attributes.

4. Discussion Urbanization is a continuous and complex spatiotemporal process of LULC change. To better understand this process, we introduced a new method within the WUDAPT framework to generate a time series of LCZ maps using a machine learning and remote sensing-based classification scheme. Our study focused on Sydney, which includes large national parks/forest areas and many of Australia’s most densely populated areas, from 1990 to 2020. This approach successfully captured the urbanization dynamics of Sydney, including urban extent, growth, and densification. The overall accuracy surpasses the requirement of the WUDAPT’s minimum accuracy threshold. The LCZ scheme has become an international standard method to explore urban morphology; a persistent challenge in this class-based classification approach is the reliability of accuracy metrics. Since the labelling of TA relies on human judgment, any mislabelling of LCZ classes can result in incorrect classifications, even when accuracy metrics appear high. For instance, some existed Sydney’s LCZ maps on the WUDAPT portal, with high OA values (>0.8), did not accurately represent the urban LULC. This indicates that a higher OA value does not always guarantee LCZ map accuracy, underscoring the importance of conducting quantitative accuracy assessments in parallel with the visual evaluation of the LCZ map. Studies have reported varying OA values for LCZ classifications globally, and this variation can be attributed to several factors. For instance, European cities have achieved higher overall accuracy values in LCZ mapping (OA > 0.9, as reported by [29]), than in Chinese cities (comparatively lower OA values) [41]. These discrepancies can be attributed to the significant differences in urban development levels between regions, which result in diverse urban structures and forms, ultimately affecting classification accuracy. Earlier LCZ classification (pre–2000) relied on Landsat 5’s Thematic Mapper, featuring a coarse thermal band and 8-bit radiometry, which supported indices like NDBAI, NDBI, EBBI, NDVI, NDWI, and BCI, but was limited by broader bandwidths and limited thermal precision [59]. After 2000, classifications transitioned to updated Landsat 7 and 8’s OLI and TIRS sensors, providing enhanced thermal resolution, and 12-bit radiometry, improving thermal-dependent indices such as NDBAI and EBBI [60]. Despite Sydney’s rapid growth and densification, the accuracy statistics for LCZ maps remain relatively consistent across all maps (OA: 0.68–0.72). This consistency can be attributed to well-labelled TAs, the robust classification scheme, and the city’s development patterns. However, a notable decrease in OAu and consequent increase in OAbu could be the influence of landscape complexity and sensor differences over time. These findings align with global trends, where the interplay of urban morphology and sensor capabilities shapes LCZ mapping accuracy. The WUDAPT framework-based single-year LCZ classification schemes have been applied in various urban studies, demonstrating good accuracy (e.g., [25,26,35,55]. However, multi-year studies, such as Demuzere et al. [54], have reported inconsistencies in tracking urban growth. While some other research incorporates additional demographic information and post-processing techniques to improve OA and reduce misclassification [41,43], these methods are often constrained by the limited availability of such data across various cities and time periods. Our TA tracking approach effectively overcame these inconsistencies. Most of the earlier studies have highlighted that well-labelled TAs are key in the classification scheme for accurately representing urban LULC [29,44,61]. A notable strength of our classification approach lies in the careful inspection and adjustment of each TA polygon across different time periods. Although this process is time-consuming, it substantially reduces the risk of misclassification in any given year, thereby enabling a more accurate time-series representation of urban LULC change (urbanization dynamics, including growth, densification, and reclassification). This approach also minimizes reliance on supplementary data, making it suitable for diverse urban areas and adaptable across different temporal scales and regions, providing a robust solution for multi-temporal urban analysis. Natural land cover classes (LCZ A–G) were classified with higher accuracy than urban built-up classes due to similar spectral reflectance characteristics within the urban classes, as well as the heterogeneous nature of built-up areas. Similar effects were reported by Bechtel et al. [15], where urban types consistently exhibited lower accuracy than natural classes. Among the built-up classes, compact mid-rise areas (LCZ 2) showed the lowest accuracy across the entire period in Sydney, likely due to the limitations of early Landsat sensors and image quality before 2000. In more recent years, lower accuracy for LCZ 2 could also be attributed to its mixed-land use nature, as compact midrise zones often contain a combination of residential, commercial, and industrial areas [54]. Although the classification scheme incorporates the most common spectral bands, the absence of building height information in Landsat images represents a significant limitation [16,54]. This shortcoming can lead to misclassification within built-up areas, particularly between compact midrise and highrise, as well as between open midrise and highrise classes [61]. Merging certain LCZ classes with similar functions (e.g., LCZ 1 and 4, LCZ 8, and 10, LCZ A and B) may reduce classification challenges by minimizing confusion between similar classes to some extent [41]. However, the decision to merge largely depends on the application of the LCZ map. For applications requiring detailed urban morphology, such as climate modeling, keeping all LCZ classes may be more beneficial, while broader urban planning might find merging similar classes more effective. Additionally, classification accuracy could be improved by using Sentinel imagery [47], but since these images are only available from 2015, this limits their utility for long-term LCZ mapping. The multi-decadal LCZ map reveals that lowrise classes (e.g., LCZ 3, LCZ 6) dominated the urban classes of Sydney, with compact lowrise areas (LCZ 3) exhibiting the highest density in the central part of the city in recent years. In contrast, open lowrise (LCZ 6) areas are more prevalent in suburban regions (Fig 3). To assess the classification approach’s effectiveness in tracking urban growth, we calculated the area covered by each LCZ class, providing insights into urban development trends. Compact and open classes (LCZ 1–5) have expanded, whereas open and sparsely built low-rise areas have significantly declined. The rapid growth of LCZ 3, particularly in southwestern Sydney (Fig 4), aligns with government initiatives that released land for housing and community development [50]. Moreover, recent infrastructure projects, such as the construction of an airport, railway lines, and roads, may have further accelerated residential expansion in these areas, particularly the LCZ 3 (captured in Fig 4c). These shifts in LCZ patterns, particularly the rise of compact built-up areas, influence urban microclimates by altering thermal environment, e.g., compact developments (LCZ1–3) often increase local temperatures, impacting public health through risks like heat stress and reducing urban resilience to climate extremes [62]. By applying LCZ analysis to these trends, Sydney’s urban evolution can be better understood in comparison to earlier mappings, such as the LCZ map by [31,51], which shares similarities with 2020 LCZ map, but excludes LCZ 5, LCZ 7, LCZ C, and LCZ E. Although LCZ 7 was omitted in our classification as well, we included the other classes despite their smaller spatial coverage, which may be attributed to the small differences in city descriptive parameter (S2 Table) values with Lipson et al. [31]. Fig 7 illustrates the land cover transformations in Sydney from 1990 to 2020, highlighting key aspects of urban change. The most notable change within the urban LCZs involves transitioning from large lowrise areas to more compact and open high- and midrise developments. Open lowrise areas have largely been converted into compact lowrise developments. Population growth in Sydney has driven urban sprawl, particularly in the city’s western suburbs, where much of the expansion has transformed sparsely built areas or agricultural land into new neighbourhoods through greenfield development [57,63]. Additionally, portions of water areas have been reclaimed for port and airport expansions and other building structures. These transformations, captured in detail by the LCZ map, are essential for urban climate modeling, as they reveal the loss of open spaces and the increasing dominance of densely built environments, both of which substantially impact local climate and environmental conditions [27,64,65]. However, some of these transitions may be partly overestimated, as the rate of pixel transitioned from its initial class to another and later reverted to its original class are high—often exceeding 50%—particularly in urban classes such as LCZ 5 (open midrise), suggesting that certain observed changes may reflect short-term fluctuations or classification uncertainty (S4 Fig). This analysis also provides a foundation for further research on the effects of greenfield development on urban climate, heat, and energy use, offering insights to inform urban planning and policy. Incorporating these dynamic transformations into urban planning and climate models is also crucial for understanding their impact on the environment and for fostering sustainable, climate-resilient cities. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 7. The chord diagram represents the transition among LCZ categories in the Sydney from 1990 to 2020. We calculated the absolute number of pixels that were LCZ X in 1990 and turned to LCZ Y in 2020. The thickness of the connecting arrow (chords) indicates the magnitude of the transitions from one class to another over time. https://doi.org/10.1371/journal.pclm.0000677.g007 A key challenge is the potential bias introduced by older Landsat images used in earlier years. These Landsat images characterized by limited spectral bands, lower calibration accuracy, and less robust atmospheric correction reduces the image clarity [66–68], affecting classification accuracy. This also limits the detection of fine-scale urban features, potentially misclassifying sparse or heterogeneous areas in Sydney’s LCZ map. To overcome these limitations, future studies should explore more advanced image tracking and detection algorithms, which could outperform the current classification methods. Enhanced techniques, such as deep learning or convolutional neural networks (CNNs), may significantly improve accuracy and sensitivity in identifying and tracking LCZ class transitions over time [46]. Adopting these advanced algorithms can substantially improve LCZ mapping accuracy, strengthening its applicability for urban sustainability assessments and facilitating more effective climate resilience strategies. Despite these limitations, the current approach remains a valid and valuable method, as it provides a systematic means for global mapping of urban areas in ways that no other dataset can achieve.

5. Conclusion This study presents a new approach to multi-decadal LCZ classification, successfully applied to the Greater Sydney Region from 1990 to 2020 at five-year intervals. Our method achieved an overall classification accuracy above 65%, effectively capturing urban growth and densification trends in Sydney. In this approach, we first generated the LCZ map for 2020, leveraging the availability of high-quality Google Street View and Google Maps imagery to validate the classifications. Using this 2020 LCZ map as a reference, we tracked each TA polygon in earlier periods, inspecting and revising them based on the land cover classes specific to each year. We also introduced new TA polygons in regions where misclassification was higher, ensuring higher accuracy across the temporal dataset. The time-series of LCZ map highlights the urban expansion and densification in Sydney. We found that the expansion of compact lowrise (LCZ 3) areas into previously open lowrise (LCZ 6) and sparsely built (LCZ 9) areas, predominantly in the southwestern Sydney. The approach also successfully identifies small-scale urban changes, illustrating its robustness in detecting detailed shifts in urban morphology. The total area of compact classes (LCZ 1–3) and open midrise and highrise classes (LCZ 4 –5) has consistently increased, while sparsely built areas have declined, and open lowrise areas have decreased after 2005. The expansion of these compact and open classes primarily reflects transitions from heavy industry, large low-rise areas, and more recently, from open lowrise and sparsely built classes. This study introduces the first multi-decadal LCZ maps for the Greater Sydney Region. The LCZ maps with urban descriptive parameters generated in this study offers valuable resource for urban planners, policymakers, and researchers, supporting sustainable urban development and enhancing urban climate modeling. Without any demographic data and additional extensive post-processing, this classification approach is readily adaptable for any city- and national-scale LCZ mapping globally, offering utility in urban simulation, planning, and environmental assessments.

Supporting information S1 Fig. Simplified schematic of the multi-decadal Local Climate Zone (LCZ) mapping workflow for Sydney (1990–2020) at 5-year intervals. For each classification year (yr), training areas (TAs) from yr + 5 were initially used and revised to reflect the land cover of yr (i.e., yr = TAs year – 5). For example, the 2015 LCZ classification used the 2020 TAs, which were subsequently modified based on 2015 land cover. The workflow involves iterative TA revision, LCZ classification using a random forest (RF) classifier, validation based on overall accuracy (OA), and final production of multi-year LCZ maps. https://doi.org/10.1371/journal.pclm.0000677.s001 (TIFF) S2 Fig. Temporal variations in Local Climate Zone (LCZ) areas within the Greater Sydney Region. The Figure illustrates the total area covered by each LCZ class over the years within the Greater Sydney Region. Some built-up LCZ classes cover very small area and are challenging to identify; therefore, the inset map presents areas smaller than 20 km². Each color represents a different LCZ class. Area calculations for each class were performed within the boundaries defined for the Greater Sydney Region as in Fig 3. https://doi.org/10.1371/journal.pclm.0000677.s002 (TIFF) S3 Fig. Transition matrix of natural Local Climate Zones (LCZ A–G) in Greater Sydney Region, at 5-year intervals from 1990 to 2020. Each pie chart represents the percentage of pixels that have transitioned from one LCZ category in 1990 to various other categories in subsequent five-year intervals and by the year 2020. The color-coded legend reflects the different LCZ categories involved in these transitions. https://doi.org/10.1371/journal.pclm.0000677.s003 (TIFF) S4 Fig. Percentage of pixels that experienced at least one reversion, i.e., transitioned from their initial Local Climate Zone (LCZ) class to another class and subsequently reverted to their original state over two periods: (a) 1990–2020 (using Landsat 5/7/8) and (b) 2000–2020 (using Landsat 7/8). The higher reversion rates in the earlier period (panel a) may partly reflect some classification uncertainty associated with pre-2000 Landsat data. https://doi.org/10.1371/journal.pclm.0000677.s004 (TIFF) S1 Table. Total Area (m²) of each Local Climate Zone (LCZ) class within the Greater Sydney Region from 1990 to 2020. The total area is computed within the defined boundaries of the Greater Sydney Region as delineated in Figure 3. LCZ 1–10 denote urban classes, while LCZ A–G categorize natural classes. https://doi.org/10.1371/journal.pclm.0000677.s005 (DOCX) S2 Table. City-descriptive parameters for Greater Sydney Region for built-up (LCZ 1–10) and natural Local Climate Zone (LCZ A–G) classes. The mean values are calculated with gridded morphology and surface cover obtained from building-resolving 3D dataset (Geoscape) for 2020 LCZ map. https://doi.org/10.1371/journal.pclm.0000677.s006 (DOCX)

Acknowledgments We gratefully acknowledge USGS and NASA for providing open access to Landsat data via Google Earth Engine. Our thanks also extend to Matthias Demuzere for the multi-temporal LCZ mapping tool. We further recognize the support and resources provided by the National Computational Infrastructure (NCI), Australia.

[END]
---
[1] Url: https://journals.plos.org/climate/article?id=10.1371/journal.pclm.0000677

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/