?
A machine learning dataset on winter roads of Krasnoyarsk Krai, Russia for the forestry and infrastructural projects
Machine learning in transport modeling has become a trend in science and industry. In this paper, we observe its main directions and focus on a dataset of seasonal road creation. Seasonality as a parameter in transport modeling has a significant impact on transport scenarios but is underestimated worldwide and in Russia, despite modern data challenges. A major challenge in seasonality studies is the absence of systematically collected data across multiple spatial scales, ranging from national to regional levels. This study presents the first machine learning–oriented dataset of winter roads for Krasnoyarsk Krai, the Siberian federal district, Russia, organized by 60 municipalities and integrating open geodata, government datasets, raster-based environmental variables, and socio-economic indicators. We propose a reproducible workflow for harmonizing heterogeneous sources, extracting geospatial predictors, and modeling winter road operational duration. Results indicate that forest area (correlation coefficient ρ=0.78), permafrost coverage (ρ=0.74), and snow depth (ρ=0.61) are the strongest predictors of winter road exploitation. Random Forest shows the best performance (R²=0.84; average exploitation forecast MAE=6.3 days) among other ML methods while linear models performed well with only two predictors. The study highlights the dominant role of natural and climatic factors in shaping seasonal road operability. It also outlines methodological directions for future regional datasets under increasing climate variability. The dataset will be extended to the entire Siberian Federal District and maintained to collect statistics for dual-season (winter and summer) accessibility modeling. This data has practical use in the forestry, road development, regional planning, and petroleum, and could interest researchers studying seasonal roads.