Answers are grounded in this paper and cite their sources.
Tomato is one of the most extensively cultivated and economically important crops worldwide, particularly in large-scale greenhouse environments, where plant health monitoring and fruit detection are essential for smart crop management. Greenhouse environments present multiple challenges: dense structures can obscure early disease symptoms and fruits, and fruits across different maturity stages and sizes need to be detected accurately. Moreover, deploying detection models on edge devices carried by mobile robotic platforms faces strict constraints on model size and computational cost. In this study, we propose a novel hybrid detection framework, DDS-Mamba-You Only Look Once (YOLO), built upon Deformable Dynamic State-space (DDS) and Mamba, a selective state space model-based YOLO to tackle the aforementioned challenges while balancing detection performance and model size. Specifically, three modules are proposed to integrate into the Mamba-YOLO backbone to enhance the performance: (1) a Spatial Deformable Channel Module (SDCM), built upon spatial adaptive downsample modules to better capture occluded and small targets; (2) a Dysample upsampling module to preserve fine-grained features adaptively, and (3) a Long Short Fusion Refiner (LSFR) module, which combines a dual-branch split Mamba block with a convolutional feature refinement module to strengthen gradient propagation and information flow. Experimental results on an image dataset collected from real greenhouse environments show that our model, with only 3.28M parameters and 8.3 GFLOPs, achieves 93.3% mAP@50-95 and 98.1% mAP@50, surpassing seven comparably-sized YOLO series models and three recent leaf disease detection models. Ablation experiments further validate the necessity of each module and its complementary interactions. Beyond training the model, we further establish a preliminary prototype that couples the DDS-Mamba-YOLO model with quadruped robot-based autonomous greenhouse inspection and Large Language Model (LLM)-based management consultation. This integrated system validates the practical deployment potential of the proposed model on edge devices and provides a preliminary blueprint for intelligent agricultural management at scale.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: