An autoencoder driven deep learning geospatial approach to flood vulnerability analysis in the upper and middle basin of river Damodar.

Flood vulnerability mapping has significantly progressed with the advent of Machine Learning (ML), bringing greater certainty to predictions. However, conventional supervised ML techniques may not be feasible in regions where recorded flood inventory data is scarce. This study introduces a novel deep learning approach using a Convolutional Neural Network (CNN)-led Autoencoder to assess flood vulnerability under such conditions. The methodology utilizes eleven causative factors, represented as geospatial layers, to characterize the regional environment. These layers are processed using CNN Autoencoder and K-means clustering to produce a flood risk zonation map for the upper and middle basins of the Damodar River. The autoencoder's reconstruction performance is evaluated using metrics Mean Squared Error (MSE), precision, recall, and accuracy apart from cluster-based indices to evaluate its classification ability. The resulting map shows that 92% of the study area is safe, while less than 8% faces moderate to very high flood risk, aligning with historical patterns and validation analysis. The study highlights the strong impact of Drainage Density on model outcomes, while certain factors like Aspect introduce noise. These findings provide valuable insights into flood vulnerability, even in data-scarce regions, aiding proactive mitigation strategies for future flood events.