A Framework for Flood Disaster Detection From Remote Sensing Images Using Spatiotemporal Fusion With Digital Twin Technology

Flood is regarded as common disaster which could cause serious devastation in any country. Typically, it is caused due to precipitation & river runoffs, specifically at the time of excessive rainfall season. Also, by the consequence of global warming and extreme environmental effects, flood becomes a serious issue to the extent on bringing negative impacts to infrastructure and mankind. The technology of sensor network has been used to monitor changes in landcovers and water level fluctuations. Moreover, effective disaster monitoring & notification system in real-time becomes a crucial part which needs to be overcome. For this reason, the proposed methodology is designed aiming at developing natural disaster prediction and monitoring system for alerting that aids in offering right decision at right time. At first, remote sensing image data are collected and preprocessed using Frequency Ratio and Multi-collinearity test (MCT) to ensure noise removal and image augmentation by enhancing their quality. A feature extraction process is carried with the use of Deep Convolution VGGNet-16 from which optimal features are selected using Improved Harris Hawks Optimization algorithm (IHHOA). Then, a Flexible Spatiotemporal image fusion (F-SPTF) approach is employed to fuse images. After this, Deep cascaded RNN classifier is employed for predicting flood occurrence and to map flood susceptibility areas. This, in turn, classifies the normal and abnormal condition of flood occurrence thus giving alerts in case of natural disaster occurrences which could be visualized through digital twin technologies. At last, the performance is carried for various metrics like F1-score, precision, accuracy, recall, RMSE, NSE, MAPE, and MAE to validate proposed scheme efficiency. A suggested scheme offers an accuracy rate of about (99.89%), precision (99.37%), recall (99.82%), and F1-score (99.74%). The error rates estimated like RMSE (0.784), MAE (0.764), and MAPE (0.102) also seems to be lower than other existing models compared.