Supervised multi-regional segmentation machine learning architecture for digital twin applications in coastal regions

Abstract

The objective of this study is to develop a global terrain and altitude map by combining a digital twin model and deep learning technique on Florida's coastal area. Utilizing USGS data, we are able to represent diverse landforms while ensuring the accuracy of elevation changes. In order to mitigate projection distortions, we rescaled 5000 map segments worldwide, ensuring that key geographical features are included. We segment the terrain into seven distinct classes: Water, Grassland, Forest, Hills, Desert, Mountain, and Tundra. The map features are enhanced by median filtering and each class is color-coded. Random parameters were introduced in overlapping image sets in order to ensure variety and prevent redundancy. On these seven terrain classes, the U-Net network is used to perform segmentation tasks. In order to monitor the performance of the model, we implemented cross-validation. The model's effectiveness is demonstrated by robust ROC curve analysis and high AUC values, which indicate accurate terrain categorization. Using deep learning methods and satellite imagery from Google Earth, the primary objective is to develop a digital twin of Florida's coastline. The digital twin serves as both a physical and simulation model, accurately resembling real-world locations. In addition to the achievement of detailed terrain mapping, this approach is likely to have significant applications in environmental monitoring and urban planning as well. In terms of reliability and performance, the digital twin model is expected to be a significant advancement in the field of geographical information systems.