Research on the Spatial Configuration and Layout of Chinese Hakka Architecture Based on the CNN-LSTM-Attention Method

This dataset comprises floor plans of 14 representative Hakka architectural sites in southern Jiangxi, China, including 6 residential walled dwellings, 7 public administrative buildings, and 1 ancient Hakka city. The plans are digitally segmented into nine directional zones according to the Later Heaven Eight Trigrams (Bagua) system. The data were used to train and validate a CNN‑LSTM‑Attention hybrid model for quantitative matching between architectural spatial layouts and traditional Chinese Five Elements and Bagua philosophy. The dataset includes image files, pixel density matrices, and extracted feature similarity scores (CNN, density, and spatial) for each directional sector. It supports research on digital heritage analysis, machine learning applications in architecture, and the quantitative understanding of Feng Shui principles in traditional Hakka construction. All 126 data points achieved a 100% directional matching rate, with an average similarity of 0.9904, demonstrating the model’s accuracy and robustness.本数据集收录中国赣南地区14处具有代表性的客家建筑遗址平面图，涵盖6座居住类围屋、7座公共行政建筑及1座客家古县城。这些平面图依据后天八卦（Later Heaven Eight Trigrams, Bagua）体系，被数字化分割为九个方位分区。本数据集被用于训练并验证CNN-LSTM-Attention混合模型，以实现建筑空间布局与中国传统五行及八卦哲学的量化匹配。数据集包含各方位分区的图像文件、像素密度矩阵，以及提取得到的CNN、密度与空间三类特征相似度评分。本数据集可支撑数字遗产分析、机器学习在建筑学领域的应用，以及对客家传统建筑中风水原则的量化认知研究。全部126个数据点均实现100%的方位匹配率，平均相似度达0.9904，验证了该模型的准确性与鲁棒性。