Hybrid Deep Learning Framework for Urban Land-Use Prediction and Scenario Modelling of Bangalore

India’s rapid urbanization in major metropolitan cities has triggered significant shifts in land-use patterns, exerting far-reaching effects on regional environmental balance and the future resilience of local ecosystems. Bangalore, as a major metropolitan, has expanded its paved surface, placing its ecological systems under significant stress. Standard methods for modelling land use often struggle with complex spatial and temporal connections. These approaches also tend to lack strength when it comes to creating forecasts based on data for extended scenarios. This research presents an innovative hybrid transformer-based framework designed to predict fine-scale urban land-use dynamics within the city of Bengaluru. The multi-temporal Land use data of Bangalore were derived from satellite image classification, alongside static and dynamic geospatial predictor variables, which were considered necessary for land use forecasting based on the literature review. The proposed model architecture is a hybrid that integrates the Convolutional Neural Networks (CNNs) for spatial feature extraction with a Transformer-based encoder, leveraging self-attention mechanisms to effectively capture complex spatio-temporal dependencies from the data.  A baseline model, using CNN encoders, has been successfully implemented and trained on the 2012-2023 dataset. Preliminary results yield a high overall accuracy and a Kappa score. The framework is designed to achieve state-of-the-art prediction accuracy by uniquely capturing both spatial and temporal dependencies. The evaluation focuses on key spatial metrics, where we project superior 'quantity' and 'allocation' agreement and a more accurate capture of the heterogeneous patterns of both 'infill' and 'expansion' growth. The validated framework will be used to simulate two critical future scenarios for Bangalore's development: a 'Business as Usual' (BAU) scenario based on historical trends and a policy-driven 'Sustainable Development' (SD) scenario. By providing geospatial forecasts of these radiating paths, this research will offer a dynamic decision-support tool, empowering planners to visualize and assess the long-term environmental and ecological impacts of future growth and to guide policy towards a more sustainable urban future.

Keywords: Deep Learning, Transformer, Convolutional Neural Networks, Classification