Intelligent housing-dimensional urban physical examination requires evaluating complex visual scenes in aging communities. Existing methods and datasets are insufficient for these heterogeneous tasks and severe class imbalances. To address this, we introduce the Housing-dimensiOnal visUal inSpection imagE Dataset (HOUSED) with a hierarchical labeling scheme, and propose a hierarchical Vision Mixture of Experts (VMoE) framework. At its core, the proposed CS-DisVMoE module utilizes a CS-Soft routing mechanism to capture spatial feature correlations, optimizing expert assignment and reducing inference overhead. Additionally, a FENNEL-based non-linear graph partitioning mechanism converts pre-trained dense weights into semantically coherent expert initializations, accelerating convergence while preserving localized visual clustering. To address the hierarchical labels, we design a composite loss function: a Supervised Contrastive Loss acts as a parent-category soft constraint to accelerate convergence, while Focal Loss mitigates data imbalance and handles fine-grained subcategory classification via hard sample mining. Across evaluated datasets, the full proposed framework improves accuracy by an average of 4.3% over the ViT-Tiny baseline and 1.81% over the best-performing VMoE baseline. Furthermore, it achieves these improvements with lower computational costs. Further tests on mixed public vision datasets verify its generalizability and competitive performance for complex-scene applications.

In recent years, the use of street-view images for urban analysis has received much attention. Despite the abundance of raw data, existing supervised learning methods heavily rely on large-scale and high-quality labels. Faced with the challenge of label scarcity in urban scene classification tasks, an innovative self-supervised learning framework, Trilateral Redundancy Reduction (Tri-ReD) is proposed. In this framework, a more restrictive loss, “trilateral loss”, is proposed. By compelling the embedding of positive samples to be highly correlated, it guides the pre-trained model to learn more essential representations without semantic labels. Furthermore, a novel data augmentation strategy, tri-branch mutually exclusive augmentation (Tri-MExA), is proposed. Its aim is to reduce the uncertainties introduced by traditional random augmentation methods. As a model pre-training method, Tri-ReD framework is architecture-agnostic, performing effectively on both CNNs and ViTs, which makes it adaptable for a wide variety of downstream tasks. In this paper, 116,491 unlabeled street-view images were used to pre-train models by Tri-ReD to obtain the general representation of urban scenes at the ground level. These pre-trained models were then fine-tuned using supervised data with semantic labels (17,600 images from BIC_GSV and 12,871 from BEAUTY) for the final classification task. Experimental results demonstrate that the proposed self-supervised pre-training method outperformed the direct supervised learning approaches for urban functional zone identification by 19% on average. It also surpassed the performance of models pre-trained on ImageNet by around 11%, achieving state-of-the-art (SOTA) results in self-supervised pre-training.