Implementing digital technology in landscape architecture: cityscape image segmentation via deep learning

Digital technology plays a crucial role in urban park landscape design by enhancing design efficiency, precision, and recall through precise data analysis, simulation, and optimization, while also aiming to improve ecological sustainability and user experience. Specifically, cityscape image segmentation, as a key digital technology, enables the precise identification and segmentation of various elements within urban street scenes, providing detailed site analysis and landscape simulation. This assists designers in making informed decisions when planning and optimizing urban layouts, resulting in aesthetically pleasing and functional environments. Therefore, we present the YOLACT-Premium model for cityscape segmentation, which includes the Enhanced Asymmetric Shuffle Network (EASNet) and the Twin-route Representation Learning Module (TRLM). EASNet enhances feature extraction capabilities, while TRLM improves the efficient fusion of multi-resolution spatial and semantic information, generating high-quality segmentation masks. Our primary experimental evaluations were conducted on a composite dataset amalgamating images from the Cityscapes benchmark and diverse field-collected urban scenes, with further robustness validation performed on the BDD100K dataset. Results demonstrate that YOLACT-Premium achieves a mean Average Precision (mAP) @0.5:0.95 of 52.05 % and a [email protected] of 75.58 %. Inference speed was benchmarked at 18.09 FPS using FP32 precision on an NVIDIA GeForce RTX 3050 GPU (8 GB VRAM) with a 640*640 input resolution and a batch size of 1. Specifically, in comparative experiments employing standard COCO evaluation protocols for object scale, wherein ’small’ objects have an area <322 pixels, ’medium’ between 322–962 pixels, and ’large’ >962 pixels, YOLACT-Premium demonstrates superior performance across these categories, achieving [email protected] values of 24.05 % (small), 48.11 % (medium), and 56.00 % (large objects), respectively. These findings indicate the model’s robust capability to handle a wide range of object sizes and its suitability for applications requiring a balance of high precision, recall, and near real-time processing speed on typical desktop-grade hardware.