<p>Rapid urbanisation in metropolitan cities of developing countries requires the immediate attention of urban planners and policymakers to ensure sustainable city development. However, many existing studies lack a comprehensive framework that integrates both multi-scale object-based classification techniques and ensemble machine learning methods to map and analyse urban land use and land cover (LULC) classification. This study fills the gap by integrating high-resolution satellite imagery, ensemble machine learning, and multi-source geospatial data to analyze urban expansion patterns in Kolkata, India. The novelty of this study lies in the integration of ensemble machine learning models for multi-decadal LULC mapping and the use of a multi-scale object-based framework to classify various land cover types across multiple hierarchical levels. This study aims to systematically investigate the Kolkata metropolitan area’s robust three-level hierarchical urban land use and land cover classification and to analyse the changing patterns of LULC with the associated urbanisation factors. The study employed machine learning (ML) and ensemble ML models for robust urban land use land cover mapping across multiple decades using high-resolution satellite imagery. The results show that the ensemble model achieves the highest accuracy, and the Level 1 classification produces the highest accuracy compared to individual ML models and the Level 3 classification. The proposed hierarchical ensemble model outperformed the best-performing individual ML models by + 2–4% (land cover) and + 2–5% (urban land-use) across 1993–2023, improving robustness of multi-decadal urban change analysis and its application to flood risk, TOD planning and climate-resilient infrastructure in the Kolkata Metropolitan Area. This study addresses the lack of an integrated framework that merges spectral, textural, and socio-economic factors with machine learning to classify multi-decadal LULC in rapidly developing Indian cities. The obtained results are highly essential for understanding robust ML and fusion model to hierarchical LULC classification in India’s metropolitan cities, which is critical for effective policy implications for sustainable urban development to support SDG 11.</p> Graphical Abstract <p></p> <p>The graphical abstract illustrates the integrated workflow adopted for hierarchical urban land use and land cover (LULC) classification through a multi-scale, object-based framework powered by ensemble machine learning (ML) techniques and multi-source geospatial data. Beginning with the acquisition of multi-sensor satellite imagery, the data is decomposed into distinct thematic layers representing number of populations, population density, non-agriculture worker, night time light data. These layers are processed and analyzed at varying spatial and spectral scales to generate enhanced composite images for detailed interpretation. Feature extraction plays a central role, enabling the transformation of raw spectral data into representative feature spaces, which are subsequently fed into machine learning algorithms. Supervised learning methods are individually and fusion of ML model was used to improve classification accuracy and robustness. The maps produced demonstrate distinct different decades LULC classification of 1993, 2003, 2013 and 2023, highlighting the machine learning and the fused ML model performance on hierarchical levels classification. Accuracy assessments confirm the reliability of the classification, while chord diagrams further emphasize the transitions between different land use categories over time. This workflow not only provides a comprehensive methodology for large-scale urban studies but also emphasizes the significance of machine learning and fusion of ML model integration with Earth observation data for monitoring urban sustainability, environmental dynamics, and policy planning across diverse geographic contexts.</p>