Answers are grounded in this paper and cite their sources.
Instance segmentation based on remote-sensing imagery provides a critical pathway for precision forestry monitoring. However, in the health management of Pinus sylvestris var. mongolica (Mongolian pine) forests, existing studies are predominantly limited to identifying either healthy or diseased trees alone, lacking a viable solution for their simultaneous and precise segmentation. This limitation hinders a comprehensive assessment of tree spatial distribution and health status, thereby constraining precise management decisions. The challenge is further exacerbated in uncrewed aerial vehicle (UAV) remote-sensing images by complex forest backgrounds, blurred adhesion between adjacent tree crowns, and an extreme class imbalance between healthy and diseased samples, collectively leading to high miss-detection rates and inaccurate segmentation boundaries. To address these issues, this article proposes a novel instance segmentation framework named BIS-Net. Our contributions are threefold: First, we construct a feature enhancement backbone by integrating the Conv2former module and adopting a global feature pyramid network-inspired feature fusion to strengthen the model’s multiscale boundary perception in complex backgrounds. Second, we propose a dynamic copy–paste data augmentation strategy that adaptively enriches minority-class samples to mitigate extreme class imbalance. Third, we develop a jointly optimized S2V loss function that couples bounding-box regression with classification weighting to improve the quality of both object detection and mask segmentation. Experiments on a self-constructed UAV Mongolian pine dataset demonstrate that our method achieves 49.1% APbbox and 38.8% APₘₐₛₖ , outperforming the baseline model YOLOv9 by 3.9% and 6.2%, respectively. Notably, the detection accuracy for diseased pines is significantly improved ( APbbox +5.9% and APₘₐₛₖ +5.2%), effectively reducing missed detections. Comparative experiments with other leading methods and extensive ablation studies further validate the effectiveness and robustness of each proposed module in addressing the challenges of remote sensing under extreme class imbalance.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: