A high-resolution (HR) image provides more detailed information about structural conditions than a low-resolution (LR) image. Increasing interest and rapid progress in Ultra-High-Resolution (UHR) segmentation have created the need for a large-scale benchmark with dense fine-grained annotations covering a wide range of scenes. The modified convolutional neural network architecture based on Global–Local Networks was used to implement the recognition algorithm. Various image segmentation applications have been found, including urban planning, forest management, climate modelling, etc. The algorithm's performance has been demonstrated on DeepGlobe and Inria Aerial datasets. The proposed method outperforms existing methods on two benchmark datasets with significant improvements.