Flood Risk Assessment using Unimodal Modality-based on Only Visual Features for Land Management

Abstract

Floods pose a serious threat to infrastructure, human livelihood and the environment worldwide. Increasing climate variability and rapid development have intensified the need for effective flood risk assessment methods. Conventional approaches generally depend on multimodal datasets comprising geospatial, hydrological and textual information, which are often constrained by data availability, processing time and infrastructure requirements. This study proposes a unimodal framework for flood risk assessment based solely on visual features derived from flood imagery. Convolutional neural networks (CNNs) including InceptionV3, VGG16 and DenseNet, were employed to classify flood conditions using images obtained from mobile phones, unmanned aerial vehicles (UAVs) and surveillance cameras. The framework addresses three assessment tasks: flood image detection, flood severity classification and postevent loss and damage assessment. InceptionV3 achieved moderate performance with F1-scores of 74%, 76.75% and 79.00% across the three tasks. VGG16 demonstrated improved accuracy and precision, particularly in severity classification (88.5%) and flood detection (83.3%). DenseNet outperformed the other models, achieving an accuracy of 92.79% and an F1-score of 92.44% in severity classification, along with strong recall (77.56%) and precision (80.33%) in post-event assessment. The results indicate that unimodal visual-based flood assessment provides a reliable and rapid solution for disaster monitoring, particularly under data limitations or rapidly evolving flood conditions.