Supervised Multi-Regional Segmentation Machine Learning Architecture for Digital Twin Applications in Coastal Regions

Abstract

This study explores the use of a digital twin model and deep learning method to build a global terrain and altitude map based on USGS information. The goal is to artistically represent various landforms while incorporating precise elevation modifications in the terrain map and encoding land height in the altitude map. A random selection of 5000 segments from the worldwide map guarantees the inclusion of significant characteristics in the subsets, with rescaling according to latitude accounting for distortions caused by map projection. The process of generating segmentation maps involves using unsupervised clustering and classification methods, segmenting the terrain into seven groups: Water, Grassland, Forest, Hills, Desert, Mountain, and Tundra. Each group is assigned a unique color, and median filtering is used to improve map characteristics. Random parameters are added to provide diversity and avoid duplication in overlapping image sets. The U-Net network is deployed for the segmentation task, with training conducted on the seven terrain classes. Cross-validation is carried out every 10 epochs to gauge the model's performance. The segmentation maps produced accurately categorize the terrain, as evidenced by the ROC curve and AUC values. The main goal of this research is to create a digital twin model of Florida's coastal area. This is achieved through the application of deep learning methods and satellite imagery from Google Earth, resulting in a detailed depiction of the coast of Florida. The digital twin acts as both a physical and a simulation model of the area, emphasizing its capability to capture and replicate real-world locations. The model effectively creates a global terrain and altitude map with precise segmentation and capture of important land features. The results confirm the effectiveness of the digital twin, especially in depicting Florida's coastline.