The study analyzes the effectiveness of semantic segmentation and instance segmentation methods for identifying anthropogenic objects with varying degrees of boundary variability in aerospace imagery. Neural network models such as U-Net, PSPNet, DeepLabv3+, SegFormer, Twins-PCPVT, ConvNeXt, YOLOv7, YOLOv8, YOLOv9, and YOLOv11 are utilized. The authors categorize object contour variability into three levels and examine its impact on model accuracy and generalizability. A key focus of the study is the relationship between contour variability and the effectiveness of deep learning approaches. The research involves annotating remote sensing data to determine the degree of boundary variability, conducting experiments with neural networks, and developing an algorithm to compare the performance of networks belonging to different segmentation types. The paper also discusses segmentation quality metrics and their interpretation nuances. The results demonstrate that semantic segmentation models are more effective for detecting large-area objects with pronounced boundary variability, while instance segmentation models achieve high recognition accuracy for objects with minimal boundary variability. In conclusion, the authors emphasize the critical role of contour variability in data preparation and segmentation method selection. They highlight the need for further research to enhance model training and improve object detection reliability.
Saidov et al. (Fri,) studied this question.