ESCA-enhanced YOLOv5s: A lightweight framework for Chinese fir seedling stage classification and quantity estimation

Due to the extensive cultivation scale of Chinese fir ( Cunninghamia lanceolata ) seedlings, accurate seedling quantification and growth stage classification have become critical yet challenging tasks in precision forestry. To address the limitation of conventional methods in accuracy and scalability, this study proposes an enhanced YOLOv5s-based detection framework incorporating three strategic improvements: (1) Shufflenetv2 backbone for model lightweighting, (2) an Enhanced Spatial-Channel Attention mechanism that improves upon Enhanced Channel Attention's capacity to resolve complex background interference and spatial dependencies, and (3) a specialized Focal-EIoU loss function for improved detection accuracy. Test results demonstrate that the improved model achieves significant improvements over the standard YOLOv5s: a 1.7 % increase in [email protected], an 88 % reduction in parameters (from 26.8 MB to 3.3 MB), and real-time inference running at 300.2 FPS. This enhanced performance incurs only a marginal 0.6 % drop in recall rate. When benchmarked against nine classic object detection algorithms, our model exhibits superior accuracy speed balance. For practical implementation, we deploy the optimized architecture via the NCNN framework, developing an Android-based application that enables field-deployable seedling counting and growth-stage analysis. This study introduces a lightweight detection model that achieves an effective balance between computational efficiency and detection accuracy in nursery environments, presenting a novel approach for large-scale assessments of seedling quantity and quality. • Efficient spatial-channel attention module integrated into YOLOv5's neck to boost target focus via enhanced channel attention. • Lightweight backbone network reduces model complexity. • Bounding box loss replaced for better regression accuracy. • Enhanced model achieves 1.7 % higher mAP, 88 % param. reduction (3.3MB), 300.2 FPS inference. • NCNN-deployed Android app for field-based seedling counting & growth classification.