Causal attention-based water depth estimation for complex flooding scenes in social media images

Social media flood imagery provides a valuable real-time information source for disaster assessment. However, existing deep learning methods often struggle to precisely focus attention on key submerged reference objects when processing such images, frequently failing to fully eliminate the influence of background clutter and other confounding factors. This compromises the accuracy of water depth predictions. This study proposes a multi-angle feature enhancement method based on causal attention, effectively reducing reliance on confounding features. Our approach, grounded in the Swin Transformer architecture, integrates causal attention to achieve feature decoupling. We further design a multi-angle causal feature enhancement structure, enabling the model to reduce its dependence on confounding features effectively. Experiments were conducted on a Chinese Sina Weibo flood dataset comprising 5,676 images and a public benchmark dataset containing 400 images. Compared to multiple classical CNNs and advanced Vision Transformer models, our proposed method achieves optimal performance across all evaluation metrics, attaining an F1 score of 72.11% and a mean absolute error of 10.87cm. By focusing on causal relationships rather than statistical correlations, this approach exhibits universality and portability when processing complex scene images.