Semantic Labeling of High Resolution Images Using EfficientUNets and Transformers

Semantic segmentation necessitates approaches that learn high-level characteristics while dealing with enormous quantities of data. Convolutional neural networks (CNNs) can unique and adaptive features to achieve this aim. However, due the large size high spatial resolution remote sensing images, these cannot efficiently analyze an entire scene. Recently, deep transformers have proven their capability record global interactions between different objects in image. In paper, we propose a new model combines convolutional transformers, show mixture local feature extraction techniques provides significant advantages segmentation. addition, proposed includes two fusion layers are designed represent multimodal inputs output network. The input layer extracts maps summarizing relationship image content elevation (DSM). uses novel multitask strategy where class labels identified using class-specific loss functions. Finally, fast-marching method is used convert unidentified closest known neighbors. Our results demonstrate improves accuracy compared state-of-the-art techniques.