Перегляд за Автор "Mazurets O."

Зараз показуємо 1 - 2 з 2
  • Ескіз
    Документ
    Neural Network Classification of Building Damage Levels From Earth Remote Sensing Data
    (2025) Ovcharuk O.; Zalutska O.; Mazurets O.; Мазурець, Олександр Вікторович
    This paper investigates neural network–based classification of building damage severity from Earth remote sensing imagery under heterogeneous acquisition conditions. To achieve scalable and reproducible building-level labels suitable for geospatial monitoring and decision-support workflows, an object-centered two-stage pipeline is adopted: buildings are first localized or segmented and then classified into an ordinal four-level damage taxonomy (no damage, minor, major, destroyed). The study leverages a benchmark-scale dataset with polygon building footprints and event-based variability, enabling rigorous evaluation with macro-averaged metrics and confusion-pattern analysis that reflects the ordinal nature of the labels. The methodology combines YOLO-family segmentation for efficient building instance extraction with high-capacity classifiers (including Vision Transformer backbones) for damage prediction from normalized building crops with controlled contextual margins. Particular emphasis is placed on representativeness and domain shift (e.g., satellite vs UAV imagery), and on evaluation protocols that reduce geographic leakage and expose generalization limits. Experimental results demonstrate strong macro-level performance (macro precision/recall/F1 ≈ 0.90+) and reliable separation of extreme damage categories, supporting the practical viability of the proposed approach for rapid damage mapping while highlighting the need for explicit domain-shift handling and calibrated uncertainty in ambiguous intermediate cases.
  • Ескіз
    Документ
    Representative Samples Forming of Urban Aerial and Satellite Imagery for Building Footprint Segmentation
    (2025) Vit R.; Molchanova M.; Mazurets O.; Мазурець, Олександр Вікторович
    This paper presents an approach to forming representative training samples of urban aerial and satellite imagery for building footprint segmentation. It is shown that the performance and generalization ability of convolutional neural networks strongly depend not only on dataset size, but also on controlled coverage of urban scene variability, imaging conditions, and annotation conventions. The proposed methodology combines large-scale satellite benchmarks with polygon footprint labels and aerial imagery from unmanned platforms as complementary domains, explicitly addressing domain shift, occlusions, and perspective distortions. Sample representativeness is assessed through the training and validation behavior of YOLO-family segmentation models, including convergence stability and metric profiles. The experimental results demonstrate stable learning dynamics and the presence of challenging boundary cases typical of real urban environments, confirming the effectiveness of the proposed data formation strategy for robust building footprint segmentation in practical geospatial applications.