Ward-Level Urban Climate Risk Intelligence Using Google Earth Engine and Hybrid Deep Learning Forecasting

Abstract

Rapid urbanization is increasing urban climate risk by raising surface temperatures, reducing vegetation-based cooling, increasing electricity and water demand, and exposing more residents to heat-related stress. This paper presents a ward-level urban climate risk intelligence framework for Bengaluru, India. The framework integrates Google Earth Engine (GEE), Landsat thermal and spectral indicators, measured ward population density, built-up measurements, deep learning-based validation, and local large language model-assisted action reporting. Instead of producing only a city-level heat map, the proposed framework generates analytical outputs for municipal planning: ward-level risk surfaces, sector-specific risk scores, model validation metrics, intervention priority ranks, and climate action reports. The GEE export contains 20,412 valid ward-month observations from 243 wards across the seven-year period 2018-2024. Each monthly record contains Land Surface Temperature (LST), Surface Urban Heat Island (SUHI), Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Built-Up Index (NDBI), population density, and built-up fraction. Five risk sectors are assessed: heatwave risk, power grid stress, water scarcity risk, urban fire risk, and public health risk. A feed-forward multilayer perceptron estimates present ward risk from current normalized indicators. In parallel, GRU, LSTM, and MLP models are trained on monthly sequences from 2018-2023 and compared for 2024 risk prediction, forming a hybrid architecture that combines transparent risk formulas with data-driven temporal validation. The present implementation identifies 17 severe-risk wards and 93 high- risk wards. The spatial neural estimator achieves a validation MAE of 0.46 risk-score units, while the temporal comparison selects LSTM as the best-performing model with a validation MAE of 1.47 and holdout MAE of 2.00. The results show that interpretable satellite-derived indicators, risk-sector formulas, and validated deep learning models can be combined to support ward-scale climate adaptation planning.