MFD-YOLO: an efficient instance segmentation model for precision monitoring of UAV-based lychee orchards.

Accurate instance segmentation of lychee canopy regions is fundamental for precision orchard management. UAV-based monitoring faces a critical dilemma: the heavy computation required for complex canopy features conflicts with the limited resources of edge devices. To resolve this accuracy-efficiency trade-off, we propose MFD-YOLO (Multi-scale-Downsampling Decoupling), a model designed for real-time UAV monitoring in orchard environments that addresses the limitations of conventional models, namely single-scale feature representation and insufficient inference efficiency. The main contributions are as follows: (1) Addressing the challenge of indistinct canopy boundaries and complex textures, we design the Multi-scale Feature Extraction (MFE) block. By employing a multi-branch parallel structure during training to capture both global contours and fine-grained leaf details, and re-parameterizing them into a single layer for inference, we enhance feature representation without incurring extra latency. (2) Addressing the issue where fine edge details are lost during standard downsampling, we introduce the Spatial-Channel Decoupled (SD) module. Unlike traditional strided convolutions that compress dimensions simultaneously, SD prioritizes channel information adaptation before spatial reduction, effectively preserving small-object features while reducing redundancy. (3) Evaluated by mAP50, precision, and recall metrics, the MFD-YOLO model performs excellently in lychee canopy segmentation tasks. It simultaneously outputs bounding box coordinates for real-time coarse localization and pixel-level masks for precise canopy delineation in complex orchard scenarios, effectively addressing practical issues in complex orchard field environments while achieving low latency on the server side. The dataset is randomly partitioned into training and validation sets at a ratio of 7:3. This model provides reliable technical support for key links including scientific pesticide application, targeted pruning, and efficient harvesting.