In the tropic, many crops are distributed in the highlands of provinces of the Andean regions at heights of 2,500 m asl and constitute the areas with the highest susceptibility to the frost events occurrence. The study objective was to propose an early frost prediction model based on the relationships between frost events and climatic variables, modeled with machine learning methods. The climatic variables were obtained from thirteen meteorological stations located inside flower crops and distributed in nine municipalities of the Cundinamarca Department. The variables registered were temperature, relative humidity, dew point, photosynthetically active radiation, and precipitation, entered as explanatory variables of frost events. The metrics used for predictive performance evaluation of the five machine learning methods examined were precision, recall, true negative rate, accuracy, and F1 score. The variables’ climatic behavior of previous hours to a frost event are low humidity, wind speed and cloudiness, and high thermal radiation. The fourth of the five trained models performed well due to their classification evaluation metrics, greater than 91%. The cross-validation and statistical analysis demonstrated the higher accuracy of the GBDT model on frost events detection.

Frost prediction in tropical high-mountain agricultural regions is difficult because sparse meteorological networks must represent strong terrain-driven microclimatic variability. This article presents a topoclimatic graph dataset for frost prediction in the Altiplano Cundiboyacense, Colombia. The released core dataset contains 23 agricultural weather stations and is seasonally focused on recurrent November–February frost periods rather than year-round continuous monitoring. It includes four consistently available meteorological variables at 30 min resolution: air temperature, relative humidity, dew point temperature, and solar radiation. The data were consolidated from multiple operational sources, harmonized to a common temporal grid, subjected to physical and consistency-based quality control, and completed through temporal and spatial reconstruction with traceability labels. The final release also provides binary frost_event and frost_warning_6h labels, point-based topographic descriptors, 1 km buffer-based raster summaries, land-cover proportions, station-level static feature vectors, and graph products including edge lists and adjacency matrices. These data products support graph-based deep learning, multimodal spatiotemporal analysis, and frost early warning experiments in a tropical mountain agroecosystem. The dataset offers a reproducible framework for integrating heterogeneous environmental observations into graph-ready representations while preserving sufficient environmental context for benchmarking frost prediction methods in data-sparse regions.