BFWSD: A lightweight algorithm for banana fusarium wilt severity detection via UAV-Based Large-Scale Monitoring

• A YOLOv8-based BFWSD model incorporating GPU-efficient Ghost (G-ghost) bottlenecks, achieving 55% parameter reduction in C2F modules and 25% smaller model size for efficient UAV deployment. • Novel Mixed Local Channel Attention (MLCA) mechanism integrated into YOLOv8′s backbone and neck, enhancing feature extraction with 2.4% precision improvement through combined local-global and channel-spatial attention. • First lightweight UAV framework for automated BFW severity detection (4 levels) and disease index calculation, demonstrating >85% accuracy at >12cm spatial resolution, for cost-effective large-scale monitoring. Banana fusarium wilt (BFW) poses a significant threat to banana cultivation worldwide, underscoring the need for timely and precise detection and monitoring for effective disease management. This study introduces an unmanned aerial vehicle (UAV)-based remote sensing methodology, utilizing a BFW severity detection (BFWSD) model to efficiently detect BFW and assess disease severity over expansive areas, thereby significantly reducing labor and time requirements. UAV-captured imagery of banana plantations was used to classify BFW severity into four distinct levels. The BFWSD model was constructed using the YOLOv8n network as the baseline. The model achieves optimization by replacing the bottleneck in the cross-scale convolutional feature fusion (C2F) module with a GPU-efficient Ghost (G-ghost) bottleneck, resulting in a 55% reduction in the parameters of the C2F module, thereby reducing memory consumption and improving detection speed. Additionally, the mixed local channel attention (MLCA) module was integrated into both the backbone and neck of the network, incorporating both local and global features along with channel and spatial information, thus enhancing the extraction of detailed leaf features and increasing detection precision by 2.4%. Experimental results demonstrate that the BFWSD model outperforms traditional object detection methods, achieving a 3.8% higher mAP50 and 6.5% higher recall than YOLOv8n. Furthermore, the efficiency of the BFWSD model in monitoring BFW was evaluated using remote sensing images at four different spatial resolutions (2.6 cm, 6 cm, 12 cm, and 25 cm). Based on the classification and localization of BFW using the BFWSD model, a distribution map of BFW was generated, alongside the calculation of the disease index (DI). Finally, to facilitate real-time field application, the BFWSD model was deployed on the Android platform, enabling seamless integration with UAV systems. Overall, this study provides a robust technical foundation for the large-scale BFW detection, offering significant improvements in monitoring accuracy and operational efficiency.