Lightweight MobileNet–YOLOv7 Fusion Model for Disease Detection in Dispersed Mountainous Farmland

山区农田中的病害识别面临巨大挑战，原因是地形破碎、照明异质、作物种类多样以及边缘设备计算能力有限。为解决这些限制，本研究开发了一种轻量化的MobileNet–YOLOv7融合架构，优化用于资源有限农业环境中的实时病害检测。拟议骨干用增强的MobileNet取代了原有的ELAN结构，该网络具备深度可分卷积和L1范数驱动的动态通道剪枝，有效保留高频病变线索，同时减少参数冗余。嵌入中深阶段的轻量级多尺度注意力模块增强对小且视觉上微妙损伤的反应，跨阶段部分融合机制在可控计算成本下选择性聚合带有更深语义的浅层细节。采用两阶段渐进训练策略，结合通过TensorRT实现的硬件感知INT8量化，使得在Jetson Nano等边缘设备上高效部署成为可能。 在公共和山地实地图像数据集结合上的广泛实验展示了该模型的优势。它实现了更优越的小目标检测精度（mAP@0.5,0.764–0.891），在不同尺度变化中表现出稳健的性能，并在强阴影、背光和遮挡条件下更为稳定。在Jetson Nano上，它以仅4.82瓦功耗和最小的温度升高，提供21.4 FPS的表现，优于主流轻量级探测器。结果表明，协调优化修剪、注意力调节和选择性跨尺度融合对于在复杂山区农业场景中实现高精度和节能的病害监测至关重要。

Disease recognition in mountainous farmlands faces significant challenges due to fragmented terrain, heterogeneous illumination, diverse crop species, and limited computational capacity of edge devices. To address these limitations, this study develops a lightweight MobileNet–YOLOv7 fusion architecture optimized for real-time disease detection in resource-constrained agricultural environments. The proposed backbone replaces the original ELAN structure with an enhanced MobileNet, which incorporates depthwise separable convolutions and L1 norm-driven dynamic channel pruning to effectively preserve high-frequency lesion cues while reducing parameter redundancy. Lightweight multi-scale attention modules embedded in the middle-deep stages enhance the response to small, visually subtle lesions. The cross-stage partial fusion mechanism selectively aggregates shallow details with deeper semantic information at a controllable computational cost. Adopting a two-stage progressive training strategy combined with hardware-aware INT8 quantization implemented via TensorRT enables efficient deployment on edge devices such as the Jetson Nano. Extensive experiments on the combined public and mountain field image datasets demonstrate the advantages of the proposed model. It achieves superior small-object detection accuracy (mAP@0.5: 0.764–0.891), exhibits robust performance across varying scale changes, and maintains greater stability under conditions of strong shadows, backlighting, and occlusion. On the Jetson Nano, it delivers a throughput of 21.4 FPS with only 4.82 W of power consumption and minimal temperature rise, outperforming mainstream lightweight detectors. The results indicate that coordinated optimized pruning, attention modulation, and selective cross-scale fusion are critical for achieving high-precision and energy-efficient disease monitoring in complex mountainous agricultural scenarios.