The pine wood nematode is responsible for pine wilt disease, which poses a significant threat to forest ecosystems worldwide. If not quickly detected and removed, the disease spreads rapidly. Advancements in UAV and image detection technologies are crucial for disease monitoring, enabling efficient and automated identification of pine wilt disease. However, challenges persist in the detection of pine wilt disease, including complex UAV imagery backgrounds, difficulty extracting subtle features, and prediction frame bias. In this study, we develop a specialized UAV remote sensing pine forest ARen dataset and introduce a novel pine wilt disease detection model, SLMW-Net. Firstly, the Self-Learning Feature Extraction Module (SFEM) is proposed, combining a convolutional operation and a learnable normalization layer, which effectively solves the problem of difficult feature extraction from pine trees in complex backgrounds and reduces the interference of irrelevant regions. Secondly, the MicroFeature Attention Mechanism (MFAM) is designed to enhance the capture of tiny features of pine trees infected by initial nematode diseases by combining Grouped Attention and Gated Feed-Forward. Then, Weighted and Linearly Scaled IoU Loss (WLIoU Loss) is introduced, which combines weight adjustment and linear stretch truncation to improve the learning strategy, enhance the model performance and generalization ability. SLMW-Net is trained on the self-built ARen dataset and compared with seven existing methods. The experimental results show that SLMW-Net outperforms all other methods, achieving an mAP@0.5 of 86.7% and an mAP@0.5:0.95 of 40.1%. Compared to the backbone model, the mAP@0.5 increased from 83.9% to 86.7%. Therefore, the proposed SLMW-Net has demonstrated strong capabilities to address three major challenges related to pine wilt disease detection, helping to protect forest health and maintain ecological balance.

Plant disease segmentation has achieved significant progress with the help of artificial intelligence. However, deploying high-accuracy segmentation models in resource-limited settings faces three key challenges, as follows: (A) Traditional dense attention mechanisms incur quadratic computational complexity growth (O($n^{2}d$)), rendering them ill-suited for low-power hardware. (B) Naturally sparse spatial distributions and large-scale variations in the lesions on leaves necessitate models that concurrently capture long-range dependencies and local details. (C) Complex backgrounds and variable lighting in field images often induce segmentation errors. To address these challenges, we propose Sparse-MoE-SAM, an efficient framework based on an enhanced Segment Anything Model (SAM). This deep learning framework integrates sparse attention mechanisms with a two-stage mixture of experts (MoE) decoder. The sparse attention dynamically activates key channels aligned with lesion sparsity patterns, reducing self-attention complexity while preserving long-range context. Stage 1 of the MoE decoder performs coarse-grained boundary localization; Stage 2 achieves fine-grained segmentation by leveraging specialized experts within the MoE, significantly enhancing edge discrimination accuracy. The expert repository—comprising standard convolutions, dilated convolutions, and depthwise separable convolutions—dynamically routes features through optimized processing paths based on input texture and lesion morphology. This enables robust segmentation across diverse leaf textures and plant developmental stages. Further, we design a sparse attention-enhanced Atrous Spatial Pyramid Pooling (ASPP) module to capture multi-scale contexts for both extensive lesions and small spots. Evaluations on three heterogeneous datasets (PlantVillage Extended, CVPPP, and our self-collected field images) show that Sparse-MoE-SAM achieves a mean Intersection-over-Union (mIoU) of 94.2%—surpassing standard SAM by 2.5 percentage points—while reducing computational costs by 23.7% compared to the original SAM baseline. The model also demonstrates balanced performance across disease classes and enhanced hardware compatibility. Our work validates that integrating sparse attention with MoE mechanisms sustains accuracy while drastically lowering computational demands, enabling the scalable deployment of plant disease segmentation models on mobile and edge devices.

Camellia oleifera leaf disease segmentation under natural field conditions is important for precision plant protection but remains challenging because lesions often show small target areas, blurred boundaries, uneven illumination, complex backgrounds, and coexisting symptoms. To address these problems, this paper proposes BFMambaNet, a Boundary-Frequency-guided Global Semantic Mamba Network for fine-grained disease segmentation. The model adopts an encoder-decoder framework and introduces a Global Semantic Mamba-based spatial selective feature modeling block to capture long-range lesion context and reduce semantic confusion. A gated wavelet spatial enhancement block is further designed to strengthen high-frequency boundary details while suppressing noisy responses. During training, boundary-frequency auxiliary supervision guides contour localization and pathological texture recovery without additional manual boundary labels. A reinforcement-learning-guided adaptive loss controller adjusts class-wise reweighting factors and loss-component weights according to the training state, improving optimization stability. A pixel-level dataset containing 1400 images and seven disease categories was constructed for evaluation. Experimental results show that BFMambaNet achieves 92.39% Precision, 91.43% Recall, 91.26% Dice, and 85.46% mIoU, outperforming representative CNN-based, Transformer-based, and Mamba-based models. Evaluations on environmental subsets confirm superior robustness, outperforming VMamba by 3.70% mIoU under uneven illumination, 3.55% mIoU under complex backgrounds, and 5.10% mIoU under coexisting symptoms. Cross-dataset validation on Apple leaf diseases further proves its generalization with 3.39% mIoU and 3.84% Dice improvements over U-Mamba, while maintaining a competitive inference speed of 30 FPS. Qualitative results also show clearer boundaries, fewer missed small lesions, and more stable predictions in complex field scenarios.