Objetivo: desarrollar una metodología para la creación de un repositorio estructurado de imágenes, etiquetas y metadatos, orientado al procesamiento digital de ortoimágenes urbanas y a la segmentación de objetos cartográficos.Metodología: se construyó un repositorio a partir de ortoimágenes de alta resolución espacial y cartografía vectorial oficial, organizadas jerárquicamente y documentados con metadatos técnicos. Se aplicaron dos enfoques de etiquetado según la morfología de los objetos, y se automatizó el proceso con Python. Para validar los conjuntos de datos, se entrenaron modelos U-Net con backbones VGG16 y DenseNet121 sobre las clases analizadas, evaluando su desempeño en términos de precisión, pérdida y calidad de segmentación. Resultados: se generron 11.771 etiquetas binarias, organizadas en seis clases cartográficas. Los modelos presentaron altos niveles de precisión en clases bien definidas, como piscinas y vías, mientras que se observó un menor desempeño en clases con alta variabilidad visual, como canchas deportivas y zonas de parqueo. Discusión: el rendimiento de los modelos estuvo condicionado por la calidad geométrica del etiquetado y la representatividad de los datos. La metodología propuesta permitió optimizar recursos y automatizar la generación de etiquetas a partir de datos abiertos. Conclusiones: el repositorio demostró ser una herramienta funcional, replicable y esca-lable para tareas de segmentación semántica, y representa un aporte al fortalecimiento de la infraestructura digital para la ciencia abierta.

Consolidated field spectral ground-truth data of urban surfaces are scarce in Latin America, where construction materials, building traditions, and aging dynamics differ from those of the temperate cities for which most existing libraries were developed. This study built and characterized a full-range (350–2500 nm) field spectral library of eight roof materials across five Colombian cities in the Andean and Caribbean regions and identified the spectral regions that maximize their separability. The sampling was designed for field spectra capture under two constraints common in these settings, the absence of prior information on material distribution and the limited accessibility of roofs, following a documented institutional protocol with structured metadata. In total, 859 signatures from 151 georeferenced roofs were acquired with a Spectral Evolution NaturaSpec spectroradiometer and a contact probe. Spectral separability was quantified with the Spectral Angle Mapper (SAM) between median endmembers, and the five materials with sufficient support were classified with Random Forest, Support Vector Machine, and XGBoost on original and Standard Normal Variate (SNV) normalized signatures, under a roof-grouped, leakage-controlled evaluation. Roof color drives the response in the visible region, whereas the shortwave infrared concentrates the absorption features that discriminate composition. The mean SAM angle between materials (23.4°) is considerably larger than that between cities for the same material (4.3–8.3°), indicating that a material's signature remains consistent across the five cities and supporting their aggregation into a single library. The library and its structured metadata provide a reusable ground-truth resource for urban Earth observation in tropical contexts.

The spectral characterization of land covers using earth observation technologies is a constantly developing and highly demanded field for the study and monitoring of natural resources, agricultural activity, and urban planning. In urban applications, these technologies are gaining ground in obtaining highly detailed primary information through field spectroscopy, with key applications in risk management, cadastral management, and climate change. This study presents the spectral characterization of eight common building roof materials, based on reflectance measurements in the 350–2500 nanometer range, for a total of 938 effective spectral signatures of eight roof materials captured across 156 sampling points in five Colombian cities during the first semester of 2025. The methodology included a specific sampling and subsampling design to capture spectral variance at the roof level, using a NaturaSpec spectroradiometer and a contact probe. An analysis of the spectral library, which included descriptive and feature importance analysis using Random Forest and Spectral Normal Variate Normalization, revealed that the material types characterized in all five cities exhibit significant similarities in their absorption patterns, trends, and peaks, with the exception of translucent materials such as plastics. Furthermore, while the spectral response is dominated by building roofs color in the visible region, the short-wave infrared region has a greater spectral separability of material type and conservation status. These spectral data and, their corresponding metadata, were captured and cataloged to contribute to the construction of a National Spectral Bank, with the goal of consolidating a comprehensive Colombian spectral database for applications in agriculture, environmental monitoring, and urban management.