CLIMATE-RESILIENCE ASSESSMENT OF INDIAN BRIDGE AND ROAD INFRASTRUCTURE: A HYBRID CNN–LSTM FRAMEWORK WITH MULTI-HAZARD VULNERABILITY MAPPING

Abstract

Rapid climate change poses an unprecedented challenge to the road and bridge network of India which comprises of more than 6.2 million kilometres of roads and 1,79,000 highway bridges. This paper proposes a novel dual-objective approach that, first, maps multi-hazard vulnerability of existing infrastructure by combining geospatial multi-criteria decision analysis (MCDA) with machine learning classifiers on top of Sentinel-2 images and, second, predicts long-term structural degradation by integrating a hybrid Convolutional Neural Network – Long Short-Term Memory (CNN–LSTM) architecture that encodes both spatial damage patterns from Sentinel-2 images and temporal degradation cycles from climate reanalysis data. The training and validation datasets are four publicly downloadable open-access benchmark datasets: (a) road damage image dataset (RDD2022) by Arya et al., (2022) (figshare DOI: 10.6084/m9.figshare.21431547) with 9665 road images from India, (b) climate projections downscaled from NEX-GDDP-CMIP6 by NASA (AWS Open Data, 1950-2100), (c) ERA5 hourly reanalysis (Copernicus CDS, 1980-2023), and (d) road/bridge network from OpenStreetMap India (Geofabrik, download.geofabrik.de/asia/india.html). The overall accuracy of the vulnerability classifier based on Random Forest (RF) is 91.4 % while the accuracy of the SVM, XGBoost and standalone CNN baselines are 85.7 %, 88.3 % and 83.1 % respectively. The CNN–LSTM degradation predictor achieves a reduction of RMSE residual service-life prediction of 2.31 years, which is 34 % better than LSTM-only models. By RCP 8.5 (SSP5), 38 % of the national highway road segments are expected to exceed critical damage thresholds before 2050 for India. The ablation study indicates that the spatial CNN branch (+4.8 %) and multi-hazard fusion (+6.2 %) are the two main contributors to accuracy improvement.