Mechanisms of complex surface background effects on UAV-based semantic segmentation stability of tobacco plants in karst mountain regions

To investigate how fragmented surface backgrounds affect the stability of tobacco semantic segmentation in karst mountainous areas, this study selected a typical karst tobacco‑growing region as the site. Sample datasets representing flat and fragmented land surfaces were constructed using UAV RGB imagery. A comparative analysis was conducted using U‑Net and DeepLabV3+. The results show: (1) Sample quantity and sample composition affected model performance in different ways. Increasing the sample size improved the recognition accuracy of both models, whereas increasing the proportion of fragmented-scene samples increased Recall but reduced Precision. (2) When the surface-background condition changed from flat to fragmented, the performance of both models declined substantially. (3) Fragmented surface-background conditions were characterized by greater textural heterogeneity and boundary discontinuity, which were associated with increased missed detections, localized false positives, target fragmentation, and boundary blurring, thereby reducing segmentation stability.