Given the rapid global urbanization and climate change, understanding the influence of urban form on urban vegetation phenology is of great importance. However, previous studies rarely focused on how the unique local climates created by urban structures impact urban vegetation phenology. Here, we investigated these effects across 934 cities in the Northern Hemisphere spanning various climate regions, using fine-resolution Local Climate Zones (LCZs) data and the VNP22Q2 phenology product. By employing a Linear Mixed-Effects Model (LMM) and implementing a strict 50% purity pixel filtering strategy to ensure the rigor of our results, our findings indicate that intense urban forms, characterized by building surface fraction and height, extend the length of the growing season (LOS) by advancing the start and delaying the end of growing seasons, despite those large-scale variations of vegetation phenology are predominantly explained by background climate. On average, the LOS for urban areas exceeds that of surrounding natural environments by 13.07 days across all studied cities. Furthermore, the LMM-based quantification reveals that the height of urban structures significantly impacts vegetation phenology more than building density, with taller types extending the LOS by 38.47 days, whereas denser types extend it by only 15.00 days. These insights reveal that urbanization contributes substantially to prolonging the urban vegetation phenological season. Highlighting the influence of urban form on vegetation phenology underscores its critical role in addressing immediate climate challenges and achieving urban sustainability. This study provides essential insights for urban planning and policy to foster ecological balance in rapidly developing urban areas.

Accurate and timely crop yield prediction is crucial for ensuring food security and maintaining stable agricultural markets, yet leveraging high-temporal-resolution, multi-source data for this purpose presents a big challenge in effectively capturing intricate variable interactions across time, especially during periods of extreme weather events. To address this issue, this study introduces an Attention and Graph Isomorphism Network-enhanced Bidirectional Long Short-Term Memory network (AGB-LSTM), which is specifically designed for estimating countylevel soybean yield in the United States. The proposed model effectively integrated diverse high-temporalresolution time-series data (5-days), including Near-Infrared Reflectance of Vegetation (NIRv), Solar-Included chlorophyll Fluorescence (SIF), and Gross Primary Productivity (GPP), along with environmental covariates. The AGB-LSTM model achieves an accuracy of R2 = 0.67 and rRMSE = 14.46 %, which outperforms traditional and advanced machine learning methods tested, such as Random Forest (RF) (R2 = 0.52, rRMSE = 17.36 %) and Transformer (R2 = 0.60, rRMSE = 15.80 %). Furthermore, sensitivity analysis revealed the capability of the model for accurate and stable yield prediction 1 to 2 months before harvest. Notably, utilizing data with finer temporal resolution (5-day and monthly composite data) continuously improves performance. Specifically, the model performance is R2 = 0.67 and rRMSE = 14.46 % for the 5-day data, and R2 = 0.55 and rRMSE = 16.81 % for the 30-day data. We also highlight the robustness of the model under extreme climate events, maintaining strong performance with R2 = 0.50 and rRMSE = 21.32 %. Finally, the validation against USDA reported yields for major North American soybean regions in 2023 show that AGB-LSTM well captures the spatial patterns of soybean yield. These findings underscore the AGB-LSTM model as a promising and effective method for yield estimation, showcasing large potential for global crop yield forecasting.

Background: Accurate and rapid crop yield estimation is essential for precision agricultural management and food security. Prior studies have focused on survey-based statistical modeling, remote sensing-based spectral-yield regressions, and simulation-informed mappers, such as the scalable Satellite-based Crop Yield Mapper (SCYM), which reduce dependence on ground samples. However, rapid yield estimation across rainfed and irrigated fields remains challenging due to the scarcity of field-level irrigation data and the incomplete representation of water supply variability. Objectives: This study aims to develop an irrigation-aware framework for rapid, province-wide 10 m winter wheat yield mapping using multi-season Leaf Area Index (LAI) and yield field experiments across multiple cities. Methods: We propose an irrigation-aware model and a data fusion framework that integrates APSIM-Wheat states, harmonic components derived from the Green Chlorophyll Vegetation Index (GCVI) and their conversion to LAI, and a random forest regressor to estimate winter wheat yield. The framework simulates watersensitive ensembles by linking crop water requirements to local precipitation scenarios, represents continuous canopy trajectories using compact harmonic fitting consistent with simulated LAI, and maps the fused features onto yield within an operational pipeline. From 2020-2023, the framework was applied in Henan Province to generate wall-to-wall 10 m yield maps. Results: The results showed that incorporating irrigation variability reduced Root Mean Square Error (RMSE) by about 30% relative to rainfed assumptions. After aggregating pixel yields to counties, the predicted yields explained 65-75% of inter-county variation, with an RMSE of 0.81-1.08 t/ha over a single year. Pooled multiyear performance remained stable. Conclusion: The proposed pipeline improves winter wheat yield estimation in irrigation-dominated agricultural systems in Henan Province by coupling APSIM-Wheat-derived mechanistic constraints with continuous remotesensing phenology features, while avoiding the need for field-specific irrigation labels or extensive site-level calibration. Implications: The framework provides a practical and computationally scalable pathway for province-wide yield mapping in regions where irrigation information is incomplete, supporting operational yield monitoring and decision-making for precision management and food-security assessment.

Drought has emerged as a critical constraint on sustainable agricultural development, particularly in water-scarce agroecosystems where multiscale hydrological stresses-such as rainfall deficits, soil moisture depletion, and groundwater exhaustion-interact to undermine crop productivity stability. However, current remote sensing frameworks lack the capacity to isolate and quantify the independent impacts of groundwater drought on crop photosynthesis and yield, leading to long-standing underestimation of deep-layer hydrological stress, especially in groundwater-dependent regions. To address this gap, we have proposed a dynamic monitoring framework based on solar-induced chlorophyll fluorescence (SIF) to assess the photosynthetic response characteristics of winter wheat across the Huang-Huai-Hai Plain under precipitation, surface moisture, and groundwater anomalies. This framework integrates dynamic time warping (DTW) for interannual phenological alignment and constructs a quantile-based dynamic baseline library that overcomes the limitations of traditional normality-based anomaly metrics. On this basis, we have developed a novel photosynthetic response anomaly index (PRAI) to characterize the spatiotemporal evolution of drought-induced photosynthetic stress. Results reveal that groundwater anomalies induce a significantly lagged crop response (mean lag approximate to +2.1 months, p < 0.01) and exert stronger influence on photosynthetic dynamics than soil surface moisture or rainfall. PRAI exhibits more concentrated and persistent negative anomalies during groundwater drought years, correlating more strongly with yield loss (R2 = 0.53) than during meteorological drought years (R2 = 0.30). Cross-validation using evapotranspiration (ET) and vegetation optical depth (VOD) further confirms PRAI reliability in capturing physiological stress. The proposed SIF-based dynamic monitoring framework not only deepens the understanding of crop eco-physiological response mechanisms to multiscale water stress, but also provides critical scientific support and methodological innovations for regional scale precision agriculture, crop model optimization, and sustainable water resource management.

Change detection from high-resolution remote sensing images lies as a cornerstone of Earth observation applications, yet its efficacy is often compromised by two critical challenges. First, false alarms are prevalent as models misinterpret radiometric variations from temporal shifts (e.g., illumination, season) as genuine changes. Second, a non-negligible semantic gap between deep abstract features and shallow detail-rich features tends to obstruct their effective fusion, culminating in poorly delineated boundaries. To step further in addressing these issues, we propose the Frequency-Spatial Synergistic Gated Network (FSG-Net), a novel paradigm that aims to systematically disentangle semantic changes from nuisance variations. Specifically, FSG-Net first operates in the frequency domain, where a Discrepancy-Aware Wavelet Interaction Module (DAWIM) adaptively mitigates pseudo-changes by discerningly processing different frequency components. Subsequently, the refined features are enhanced in the spatial domain by a Synergistic Temporal-Spatial Attention Module (STSAM), which amplifies the saliency of genuine change regions. To finally bridge the semantic gap, a Lightweight Gated Fusion Unit (LGFU) leverages high-level semantics to selectively gate and integrate crucial details from shallow layers. Comprehensive experiments on the CDD, GZ-CD, and LEVIR-CD benchmarks validate the superiority of FSG-Net, establishing a new state-of-the-art with F1-scores of 94.16%, 89.51%, and 91.27%, respectively. The code will be made available at https://github.com/zxXie-Air/FSG-Net after a possible publication.

Accurate crop phenology information is essential for optimizing agricultural management and understanding agroecosystem responses to climate change. However, existing datasets are usually constrained by limited spatial and temporal resolutions and by oversimplified phenological stages. In this study, we proposed a stepwise strategy for fine-scale crop phenology monitoring by integrating multisource remote sensing data. First, we generated dense time-series remote sensing data at 30-m spatial resolution using a quick spatiotemporal fusion (STF) approach with coarse- and fine-resolution scale transformation errors (STEs) and a pixel-based synthesis base image pair (STEPSBI), enhancing temporal continuity while preserving spatial detail. Then, we assessed the separability of six phenological stages, based on the Gaussian probability density distributions of candidate features, which enabled the selection of optimal input combinations. Finally, we utilized random forest (RF) regression to model vegetative and reproductive growth stages, generating the spring maize phenology from 2001 to 2020 in Northeast China. Results indicated that the proposed stepwise strategy outperformed 1-D U-Net, phenology Seq2Seq network (PSeqNet), and improved shape model fitting models, with root-mean-square error (RMSE) ranging from 5.29 to 9.45 days in six phenological stages. Furthermore, compared with models using only spectral reflectance, vegetation indices, or temperature-derived factors, our approach reduced total RMSE by 15.70, 20.33, and 23.13 days, respectively. Moreover, the STEPSBI-fused images exhibited high agreement with the raw imagery, with the correlation coefficients (R) of the bands ranging from 0.81 to 0.88, effectively overcoming the spatial and temporal resolution limitations of remote sensing images. The generated dataset is consistent with existing phenology products from Landsat (30 m), Moderate Resolution Imaging Spectroradiometer (MODIS) (500 m), and GLASS (1 km), while providing finer spatial details and capturing interannual variability. These advantages support precision agricultural management and improve our understanding of the agroecosystem response to regional climate change.

China is one of the most biodiversity-rich countries, yet the ecological gap between its eastern and western regions, driven by geographical barriers, restricts species migration and disrupts ecosystem connectivity. However, the potential of biological flow to bridge this divide remains poorly understood. To address this, we developed a dynamic biological flow framework, combining ecological networks with two new tools, Ecological Linkage Tool Direction and Biological Flow, to quantify species migration directions and intensities across the Middle Spine of Beautiful China. Within and beyond ecological sources, we analyze dynamic biological flow and network structure at node, link, and graph levels. We also simulate changes in network efficiency under various corridor and habitat degradation scenarios. Among the 13,800 ecological corridors, 27 % are oriented east-west (EW), yet the region exhibits a net negative habitat inflow, with 48 % of migrating species potentially failing to reach target habitats. Biological flows outside habitats follow west-to-east (21 %) and north-to-south (18 %) patterns, with the highest migration losses occurring in the north-to-south direction. The regional network efficiency is 0.053. The failure of ecological network in the Inner Mongolia Pastoral Area reduces efficiency by 24 %, intra-patch EW corridors by 57 %, and inter-patch west-to-east corridors by 13 %. Species-specific analyses of the red panda (Ailurus fulgens) and Chinese horseshoe bat (Rhinolophus sinicus) reveals that habitat distribution determines dynamic flow direction, while species-specific adaptability influences flow intensity. This study quantifies dynamic biological flow patterns, overcoming the limitations of widely used static distribution-based conservation planning, more accurately reflecting species migration traits.

Observing greenspaces from inside buildings benefits urban residents' mental health. With urbanization increasingly favoring vertical growth, assessing the green visual accessibility in three dimensions (3D) is critical. Traditional Viewshed-based methods suffer from data availability and low computational efficiency, hindering large-scale, floor-specific green visual exposure analysis at the building level. Here, we developed an efficient empirical method to calculate the Green View Index (GVI) for each floor of urban buildings. Using the latest multi-source high-resolution geospatial data, we first generated a dataset of actual green visibility via the Viewshed model. We then estimated the GVI using Random Forest models featured by building and greenery metrics. Numerical and spatial evaluations confirmed good agreement with Viewshed-derived results (R2 > 0.60), enabling large-scale 3D GVI analysis. In Beijing, our model results reveal spatial and height-related disparities in GVI, with floors 4-6 showing the highest green visibility on average, given that large patches of clustered greenery are surrounded by mid- to low-rise buildings. This suggests Beijing should enhance green visual benefits for high-rise residents. The proposed workflow only requires commonly available 3D built-form and vegetation datasets, and hence it can be readily transferred to other cities after local calibration. It provides a scalable tool for quantifying 3D green visual exposure, offering valuable applications in urban planning, socio-ecological studies, and epidemiology.

Nighttime light (NTL) data serve as critical indicators of human activities and have been widely applied in urbanization monitoring and socioeconomic analyses. The two most widely used global NTL datasets, derived from the Defense Meteorological Satellite Program Operational Linescan System (DMSP-OLS) and the Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) aboard the Suomi National Polar-orbiting Partnership satellite, differ substantially in spatial resolution and temporal coverage, which hinders their direct integration into a consistent long-term dataset. Previous studies have explored the construction of annual or aggregated NTL data, but these methods often smooth out short-term fluctuations and seasonal variations. Monthly NTL, on the other hand, can provide a more detailed representation of temporal variations. However, the challenge with monthly data lies in maintaining consistent spatial resolution while capturing high-frequency temporal variations tied to economic cycles and seasonal trends, with data gaps persisting, further complicating the generation of continuous, high-resolution monthly NTL datasets. To overcome these challenges, we propose a super-resolution network for DMSP reconstruction, with dedicated pre-and post-processing to generate long-term monthly VIIRS-like NTL products (MVNL). Leveraging multi-modal observations, monthly VIIRS-like products are reconstructed by translating calibrated DMSP data from 1992 to 2013, with 2012 and 2013 serving as the overlapping years between the DMSP and NPP-VIIRS datasets. In particular, the 2013 annual data were used for model training and cross-sensor mapping, and the 2012 monthly NPP-VIIRS data were used as an independent validation benchmark. To construct the long-term VIIRS-like time series, we additionally gap-filled missing observations in the monthly NPP-VIIRS data for 2012-2024 and performed temporal correction to the reconstructed 1992-2012 NTL using the monthly NPP-VIIRS data from 2012 to 2013. Compared with the VIIRS NTL of Earth Observation Group (EOG), the extended dataset shows substantial agreement during the overlapping months in 2012, with a mean R2 of 0.65 and RMSE of 14.27 at the pixel scale and an even higher mean R-2 of 0.96 at the city scale, underscoring the reliability of the reconstructed dataset for city-level applications. The 2012 annual composite derived from MVNL shows strong agreement with the EOG product, with R(2 )values of 0.72 at the pixel scale and 0.98 at the city scale. Moreover, city-level evaluation against radiance-calibrated DMSP products further verifies the reconstruction accuracy, with an R-2 exceeding 0.94. Compared with existing NTL datasets, MVNL achieves substantial improvements in resolution, spatial calibration accuracy, and temporal continuity, establishing a continuous and trustworthy data resource. The extended monthly VIIRS-like NTL dataset for 1992-2024 is freely available online at https://doi.org/10.25442/hku.31321315.v2 (Cheng et al., 2026a).

Mapping corn distribution is challenging when samples are sparse and optical data is frequently cloudcontaminated. We propose a lightweight sample generation strategy that links peak signals with whole-season consistency to derive high-confidence corn samples. It includes three components: (1) aligning corn peaks within a phenology-aware time window, (2) integrating corn separability with two complementary indices-the corn spectral index, combining red-edge and shortwave infrared signals, and the corn radar index, based on Sentinel-1 polarization channels to ensure structural-dielectric consistency; and (3) enforcing whole-season similarity in timing, shape, and amplitude using a novel Peak- and Temporal-Weighted Dynamic Time Warping (PTW-DTW) algorithm. Compared to existing algorithms, PTW-DTW improves class separability, especially for non-corn crops with similar peaks but different shoulders. Experiments across three sites achieved overall accuracies of 91.40%, 91.70%, and 95.64%, outperforming single-source baselines and remaining stable across years and regions. County-level area estimates correlated well with official statistics (R-2 > 0.98). The generated samples matched well with field data, achieving robust accuracy with fluctuations below 0.5%. Mixed them increased accuracy by about 2%. Relying only on standard composite and common input, this strategy provides consistent regional corn maps and transferable samples in data-limited settings.

Understanding the dynamics of global built-up heights is crucial for fostering sustainable urban development and to manage urban environments effectively. While prior studies have leveraged satellite data to map built-up heights, study exploring their dynamics over extended time frames at a global scale is limited. To address this gap, we utilized radar time series data to analyze global built-up height dynamics. Initially, we calibrated radar data from multiple sources to prepare a temporally consistent dataset. Subsequently, we estimated annual built-up heights and volumes from 1995 to 2018 at a global scale (5.5-km resolution) using an enhanced built-up height estimation model. Our results indicated that the estimated built-up heights, ranging from 0 to 6.5 m, closely correlate with reference datasets, demonstrating a root mean square error of 0.32 m and an R2 value of 0.69. By 2018, the global built-up volume had reached 931.17 km3, nearly 2.4 times greater than 273.09 km3 of 1995. Notably, the relative growth rate of built-up volume in the Global South is approximately double that of the Global North. By shifting the focus from traditional horizontal urban expansion to vertical built-up volume growth, this study provided deeper insights into urban evolution. This approach not only aids in the development of urban growth models but also enhances research on energy consumption and carbon emissions.

Crop type mapping is a core topic in agricultural remote sensing, playing a strategic role in global food security and resource management. Advances in remote sensing and artificial intelligence (AI) have shifted crop type mapping from traditional expert-driven approaches toward data- and knowledge-driven paradigms. However, a systematic synthesis that links key components of crop identification across data sources, methods, and application contexts remains limited. In this review, we analyze the evolution of crop type mapping over the past five decades through a large-scale meta-analysis of more than 19,000 publications retrieved from the Web of Science (WoS) Core Collection. The literature was examined using a large language model-assisted workflow applied to titles, abstracts, keywords, and publication metadata, enabling scalable identification of thematic patterns and methodological trends. Building on this analysis, we organize existing studies within an analytical framework that connects crop sampling strategies, feature engineering, algorithm architectures, and validation practices. The review critically assesses empirical approaches, machine learning (ML) and deep learning (DL) algorithms, transfer learning strategies, and hybrid modeling frameworks, highlighting recent progress in deep feature extraction, learning under limited data conditions, and modeling in complex agricultural environments. Rather than proposing a universal solution, this review provides structured methodological guidance by clarifying the applicability, strengths, and limitations of different approaches under varying environmental conditions, data availability, and mapping objectives.

Accurate and scalable maize mapping is essential for reliable yield prediction and efficient resource allocation. However, many existing approaches rely heavily on large training datasets and often underuse prior knowledge, which limits their performance in data-scarce regions and weakens spatiotemporal transferability. To address these challenges, we developed a Height-Spectral Gaussian Mixture Model (HSGMM) that integrates maize canopy relative height indicators from the Global Ecosystem Dynamics Investigation (GEDI) lidar shots and the plant nitrogen status indices derived from Sentinel-2 imagery. We propose a novel height label to accurately indicate crop relative heights and achieve robust spatial extrapolation across six different test sites. On the spectral side, we adapt the Dual-Peak Canopy Nitrogen Index (DCNI) to Sentinel-2 bands and combine it with the Red Edge Position to construct a composite maize separability index, termed the DCNI-REP index (DRI). HSGMM further incorporates an adaptive penalty with an optimized Bhattacharyya coefficient ratio to tighten class separability in the joint height and spectral feature space, thereby reducing confusion with spectrally similar crops such as soybean and sorghum. Across six test sites and three years, cross-validation results show that the HSGMM model achieves overall accuracies of 0.87 to 0.95 and F1 scores of 0.86 to 0.95, consistently outperforming random forest classifiers, which achieve overall accuracies of 0.79 to 0.89 and F1 scores of 0.80 to 0.89, as well as other HSGMM variants. With a lightweight and modular design and minimal training sample requirements, our HSGMM model offers a practical solution for regional maize mapping and annual crop distribution updates in data-poor environments.

Long-term building height data are critical for analyzing urban morphological evolution and renewal processes, yet such datasets at fine spatial resolutions remain scarce for large geographical regions. This study proposes a framework to generate continuous annual building height maps for China at 30 m spatial resolution from 1990 to 2019, integrating multi-source remote sensing data (Landsat, Sentinel-1/2, etc.) through the eXtreme Gradient Boosting (XGBoost) model. The framework reconstructs Vertical-Vertical (VV) band, incorporates reference data derived from the Continuous Change Detection and Classification (CCDC) algorithm, and utilizes Total Variation (TV) denoising to achieve temporal consistency, while retaining inter-annual building height variations. Validation results demonstrate the stable performance of the building height maps over the past three decades, with nationwide RMSE values ranging between 3.20 and 4.27 m. Comparisons confirm that our results are consistent with reference height datasets and capture the temporal evolution of building heights driven by urban development and renewal. Furthermore, our dataset shows pronounced horizontal and vertical expansion of Chinese cities between 1990 and 2019, as the total impervious surface area increases from 54 407.03 to 170 595.06 km2 and overall building volume rises from 365.98 to 882.43 km3. Provincial contributions to national building volume change substantially over time, with Hebei (11.3 %), Shandong (10.6 %), and Heilongjiang (10.2 %) leading in 1990, while Shandong (9.8 %), Hebei (9.5 %) and Guangdong (9.4 %) are in the leading positions in 2019. The resulting annual 30 m resolution building height datasets, made openly accessible, provide a valuable foundation for cross-city comparisons, long-term three-dimensional (3D) urban morphology studies, and policy-relevant planning in fast-growing Chinese cities. The dataset is available at https://doi.org/10.6084/m9.figshare.29918978 (Zhang et al., 2025b).

Abstract. Nighttime light (NTL) data serve as critical indicators of human activities and have been widely applied in urbanization monitoring and socioeconomic analyses. While the most utilized global NTL datasets are derived from the Defense Meteorological Satellite Program Operational Linescan System (DMSP-OLS) and the Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) aboard the Suomi National Polar-orbiting Partnership satellite, the inherent differences in spatial resolution and temporal coverage between these sensors present challenges for direct integration into a consistent long- term dataset. Previous studies have explored the construction of annual or aggregated NTL data, but these methods often smooth out short-term fluctuations and seasonal variations, limiting the ability to capture fine-scale temporal dynamics. Monthly NTL, on the other hand, can provide a more detailed and accurate representation of temporal variations. However, the challenge with monthly data lies in maintaining consistent spatial resolution while capturing high-frequency temporal variations tied to economic cycles and seasonal trends, with data gaps persisting, further complicating the generation of continuous, high-resolution monthly NTL datasets. To bridge this gap, we propose a super-resolution network for DMSP reconstruction, with dedicated pre- and post-processing to generate long-term monthly VIIRS-like NTL products (MVNL). Leveraging multi-modal observations, monthly VIIRS-like products are reconstructed by translating DMSP data from 1992 to 2013 using NPP-VIIRS data from 2013 to 2024 as the reference. Compared with the VIIRS NTL of Earth Observation Group (EOG), the extended dataset shows substantial agreement during the overlapping months in 2012, with a mean R2 of 0.65 and RMSE of 14.27 at the pixel scale and an even higher mean R2 of 0.96 at the city scale, underscoring the reliability of the reconstructed dataset for city-level applications. The 2012 annual composite derived from monthly data shows strong agreement with the EOG product, with R2 values of 0.72 at the pixel scale and 0.98 at the city scale. Moreover, city-level evaluation against radiance-calibrated DMSP products further verifies the reconstruction accuracy, with an R2 exceeding 0.94. Compared with existing NTL products, our dataset achieves substantial improvements in resolution, spatial calibration accuracy, and temporal continuity, establishing a continuous and trustworthy data resource. The extended monthly VIIRS-like NTL dataset for 1992–2024 is freely available online at https://doi.org/10.25442/hku.31321315.v2 (Cheng et al., 2026).

Monitoring urban form change is critical for urban planning and management. Studies on three-dimensional (3D) urban form change are often constrained by coarse resolution (e.g., 5 km), limited temporal coverage (e.g., 3 to 4 years), and small sample size (e.g., less than 10 cities), limiting their capacity to capture fine-scale variations across urban landscapes. This is especially relevant to China, a country characterized by large-scale urbanization, and heterogenous urbanizing pathways. Here, we studied typologies of 3D urban form change across 305 Chinese cities between 2015 and 2024 at 100-m spatial resolution and analysed their distributional patterns. Through unsupervised classification on Sentinel-1 Synthetic Aperture Radar backscatter time series data, we identified four types of 3D urban form change: upward growth (2.20% of the total area), demolition (4.36%), fluctuation (8.60%) and stable (84.84%), with an overall accuracy of 81.09%. Upward growth was more common in southwestern and southern provinces, which typically have medium Gross Domestic Product (GDP) levels and low urbanization rates. Demolition was more prominent in central and western provinces, and in provinces with low GDP and low urbanization rates. Fluctuation was often found in northwestern provinces where GDP and urbanization rates tend to be lower. In contrast, the stable type dominated southern coastal and northeast provinces characterized by higher GDP and urbanization rates. Through this effort, we captured nuanced types of urban form change such as demolition and fluctuation. The proposed method can also be applied to ongoing planning practices, such as spatial planning monitoring and urban renewal, in China.