Pine wilt disease (PWD), caused by the pine wood nematode (Bursaphelenchus xylophilus), continues to threaten forest ecological security. Unmanned aerial vehicle (UAV) remote sensing makes large-scale screening feasible, but accurate canopy segmentation remains difficult in practice. The main obstacles are threefold: (ⅰ) canopy appearance changes noticeably from early to late infection stages, which can cause model representations to drift toward later-stage symptoms and weaken subtle early-stage cues; (ⅱ) dense pixel-wise annotation is costly, making semi-supervised learning dependent on imperfect pseudo-labels; and (ⅲ) forest backgrounds are cluttered and often visually similar to diseased regions, limiting the discriminative ability of RGB appearance alone. To address these three practical difficulties, we build a standardized UAV canopy dataset for PWD and develop a lightweight multimodal segmentation framework. The method combines three components. First, Nested-Tempo Memory Consolidation (NTMC) is designed as a stage-aware extension of EMA-based teacher–student learning. It maintains nested fast/medium/slow temporal trajectories to retain stage-specific knowledge while improving cross-stage stability during Early–Middle–Late sequential training. Second, Drift-Compensated Consistency Regularization (DCCR) integrates reliability calibration into semi-supervised consistency learning, so that unlabeled samples contribute training signals mainly when they are sufficiently reliable, reducing error accumulation from noisy pseudo-supervision. Third, Vegetation-index-conditioned Cross-Modal Attention (VCCA) uses vegetation indices—Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI)—as physiological cues to modulate visual features, thereby reducing the dependence on RGB appearance and improving feature discrimination under texture-similar forest backgrounds. Experiments on one in-house dataset and three external datasets show consistent improvements in mean intersection over union (mIoU), F1-score, and Matthews correlation coefficient (MCC). With all components enabled, the framework improves mIoU from 0.6015 to 0.6829 over the baseline and produces cleaner disease boundaries with fewer background false alarms, demonstrating its potential for practical UAV-based forest disease monitoring.

Accurate individual-tree segmentation in rubber tree plantations is essential for resource inventory and agroforestry management, but it remains difficult in densely planted stands because adjacent crowns often overlap, point density varies markedly along the vertical canopy structure, and tree morphology differs across growth stages. To address these challenges, we introduce MDA-SegNet, a multimodal point-cloud instance segmentation framework designed for densely planted rubber plantations. The Multimodal Deformable Encoding (MDE) module integrates top-view orthophoto cues with light detection and ranging (LiDAR) geometry, providing complementary crown-boundary information for more consistent separation of adjacent trees. The Z-Order Selective Mamba (ZOS-Mamba) module converts three-dimensional spatial structures into locality-preserving sequences and models long-range vertical dependencies under uneven point-density distributions. The Adaptive Lemming Optimization Clustering (ALOC) module further refines instance separation by adapting to canopy overlap and morphological heterogeneity. Experiments were conducted on our self-constructed rubber tree dataset (RT-Set) and five public forest datasets. Compared with current state-of-the-art models, MDA-SegNet achieved higher F-score and mean Intersection over Union (mIoU), showing stronger robustness and cross-domain generalizability. On RT-Set, it obtained an F-score of 87.32% and a Recall of 82.90%, with improvements of 6.86 and 7.24 percentage points, respectively. These results indicate that MDA-SegNet can provide reliable individual-tree segmentation for dense rubber plantations.

As an important tropical cash crop, rubber trees play a key role in the rubber industry and ecosystem. However, a significant challenge in precision agriculture and refined management of rubber plantation lies in the limitations of traditional point cloud segmentation methods, which struggle to accurately extract structural parameters and capture the spatial layout of individual rubber trees. Therefore, we propose an optimized dual-channel clustering method for the UAV LiDAR-based Rubber Tree Point Cloud Segmentation Network (RsegNet) for improved assessment of rubber tree architecture and traits. Firstly, we designed a cosine feature extraction network, termed CosineU-Net, to address the branch-and-leaf overlap problem by calculating the cosine similarity of the spatial and positional features of each point, leveraging deep learning approaches to improve feature representation. Secondly, we constructed a dual-channel clustering module reducing prediction error in rubber tree point cloud data, integrating multi-class association and background classification to tackle background interference. The cluster identification and separation accuracy in high-dimensional data processing is enhanced through a dynamic clustering optimization algorithm. In our self-built dataset and across five regions of the FOR-instance forest dataset, RsegNet achieved the best performance compared to five state-of-the-art networks, reaching an F-score of 86.1%. This method calculated structural attributes including height, crown diameter, and volume for rubber trees in three areas under different environments in Danzhou City, Hainan Province, providing robust support for precise monitoring, plantation management, and health assessment.