Simulation of Natural Landscapes Using Deep Learning and VR Rendering Models in Digital Art for Modern Historic Buildings

Abstract

This technical reappraisal reconstructs a two-stage deep-learning workflow for semantic analysis and visual rendering of landscape architectural plans. The intended pipeline uses a fully convolutional network (FCN) to assign labelled pixels to 11 landscape categories and then applies a cycle-consistent generative adversarial network (CycleGAN) to semantic-map patches. The surviving inventory records 328 plan images and 57,271 connected annotated elements. An unaudited internal-validation summary records pixel accuracy above 92\% and mean Intersection over Union of 0.87; these figures are historical descriptors rather than independently verifiable performance results because the source-level partitions, per-class scores, repeated runs, code, and external test set are unavailable. Displayed CycleGAN checkpoints from Epochs 50–300 show greater texture density but persistent structural distortions. The surviving examples use prepared semantic masks, and no retained output verifies the complete FCN-prediction-to-CycleGAN pathway. The reappraisal therefore documents the intended integration, corrects an algebraically redundant mask-summation formulation, identifies a clean-mask-to-predicted-mask domain shift, and delineates the evidence required for reproducible evaluation. The record supports feasibility of the stages as an archived demonstration, but not reproducibility, end-to-end validation, or superiority over alternative methods.