The aim of this work is to develop an automated pipeline that extracts solar direc- tional information of the buildings (i.e., the orientation of windows, rooms, walls and balconies) from cadastral scans. This work is going to support the design of energy efficiency policy in Italian residential public housing, a setting where public resources are typically scarce and tenants suffer from different vulnerabilities.

Specifically, in the GRINS project, Spoke 6, work package 2, the research line on energy poverty has developed two pilot surveys to collect information from public housing tenants in Reggio Emilia and in Foggia on the energy efficiency of their homes, their energy expenditure, and their perception of thermal comfort.

The present work will com- plement the surveys’ outcomes by adding relevant information from cadastral scans. This information has to be structured, so the needed energy-environmental anal- yses to be developed by merging different databases, can deliver key elements to design targeted and effective actions. By applying Artificial Intelligence (AI) methods (i.e. models that solve detec- tion and classification tasks), the model generates a dataset (CSV/Excel) suitable for analyses of sunlight exposure. The model to extract information is designed to minimize manual labeling, preserve anonymity, and enable scalability for large volumes of incoming scans.

Given the approach adopted to extract information, the present model can be usefully adapted to complement information in settings other than public residential housing, where policies to reduce carbon emissions from housing need to be planned. From a methodological standpoint, the present model builds on state-of-the-art computer vision techniques. The YOLO (You Only Look Once), namely the object detection architecture, is widely recognized for its speed and accuracy, and has been successfully applied in domains such as autonomous driving, surveillance, and document image analysis.

Similarly, ResNet architectures represent a milestone in deep convolutional networks, providing strong performance in visual classification and orientation tasks. In document understanding, deep learning approaches such as LayoutLM and DiT have demonstrated effectiveness in parsing complex layouts, though most focus on text, tables, or diagrams rather than architectural scans. The novelty of the current model lies in both its coding integration and appli- cation. Technically, it unifies object detection (YOLO) and orientation classifica- 2 tion (ResNet18) into a single end-to-end pipeline tailored to cadastral scans, going beyond conventional document parsing by coupling layout detection with compass- based directional mapping.

From an application perspective, it offers an automated solution for transforming raw cadastral data into machine-readable datasets, enabling scalable and anonymous analysis in support of sustainable urban policies, among other potential applications. To the best of our knowledge, up to now, the leveraging with geospatial data has not been widely explored in the literature and it can thus represent a completely new field for further investigations and related applications.

This combination of technical rigor and practical relevance underlines the novelty of the current model in the fields of document understanding and applied computer vision. The paper is organised as follows: in Section 2 we describe existed literature that can be used to solve our task ; in Section 3, we discus what methods and why we choose them to solve the problem; in Section 4 we demonstrate results and have some discussion on them with ideas for the future; Finally, Section 5 concludes all paper.