研究前沿 · RESEARCH

边缘AI与物联网在作物健康监测中的部署性能

聚焦边缘计算与物联网架构在无人机、田间节点上实时识别作物健康与病害的准确率、延迟与部署可行性,比较不同轻量模型与传感融合方案。

4 篇文章3 组独立证据2026-11-14 — 2026-11-17内容更新 2026-11-19

专题速览

先了解整体进展

本专题汇总四项来源:一篇综述、两篇摘要级系统论文和一篇全文无人机框架论文,各来源任务、数据与验证条件不同,结果只在各自条件下成立。综述提出五层边缘架构,判断病害CNN与土壤湿度预测成熟度高、养分模型依赖传感器精度、强化学习灌溉仍属新兴方向,并列出成本、数据稀缺与连通性等挑战,但未开展原始实验。摘要级研究称,树莓派加Coral Edge TPU的量化模型达到94.82%准确率,推理延迟约38毫秒、吞吐26.3张/秒,并据示例性研究估计可节省22%灌溉用水,摘要未提供测试集规模与作物构成。无人机全文研究在1260张留出真实航拍影像上取得健康、受胁迫、患病三类98.2%、92.5%、95.8%的准确率,但端到端流程仅在AirSim仿真中验证,62 FPS为RTX 3060代理基准,物理Jetson NX与田间验证尚未完成。Arduino气象站摘要仅称传感器读数准确、数据处理可靠,未报告精度、延迟或节水数值,不提供可量化的部署性能数据。[1][2][3][4]

本次更新建立专题,整理关联文章与研究背景

关键发现

共同之处与差异都保留
单项发现

综述将边缘部署的病害诊断、养分分类与灌溉决策整合为五层架构,并判断各组件成熟度不一:病害CNN与土壤湿度预测成熟度高,养分模型受低成本传感器精度制约,强化学习灌溉仍属新兴方向、需谨慎验证。该判断来自综述对既有文献的归纳,并非本综述自建实验的结果。[1]

单项发现

据该摘要,树莓派结合Coral Edge TPU部署量化模型后,报告准确率94.82%、精确率94.47%、召回率94.05%、F1为94.26%,推理延迟约38毫秒、吞吐26.3张/秒,并降低CPU占用;摘要未说明测试集规模与作物类别构成,上述数值仅在该摘要所述实验条件下成立。[2]

单项发现

无人机研究在来自卡纳塔克邦三个农场的8400张真实航拍图像上微调MobileNetV3-Large,并在1260张留出真实影像上评估,健康、受胁迫、患病三类准确率分别为98.2%、92.5%、95.8%;RPLiDAR株高与冠层体积估计相对模拟真值的平均绝对误差为14.3厘米。该分类任务为椰子棕榈冠层健康三分类,与上述树莓派研究的作物状态分类任务不同,数值不可直接比较。[4]

单项发现

该无人机框架的端到端流程(推理、LiDAR处理、自主覆盖、仪表板与报告)仅在Microsoft AirSim仿真中经五次任务验证,平均18.4分钟完成全田覆盖;62 FPS是在RTX 3060 GPU上作为Jetson NX代理测得,物理硬件基准、热特性与田间验证尚未完成,因此其延迟与吞吐不能视为机载实测结果。[4]

单项发现

Arduino Uno气象站摘要称系统可连续监测温度、湿度、气压、降雨与土壤湿度,实验测试显示传感器读数准确、数据处理可靠,适用于中小规模应用;但摘要未报告具体精度、延迟或节水数值,也未涉及边缘AI推理,因此该来源只能作为低成本田间传感的背景,不能支撑本专题关于准确率、延迟或部署可行性的量化结论。[3]

研究比较

对照研究条件与局限,具体指标请结合原文阅读。

研究对象、方法、数据条件、结果及局限
研究对象方法数据条件结果局限
Ai-Based Smart Agriculture: A Review Of Edge Disease Diagnosis, Soil Nutrient Profiling And Reinforcement Learning For Precision Irrigation

[1][1]

边缘智能架构中的病害诊断、养分分类与灌溉决策组件(综述,无原始实验)文献综述,提出五层架构(田间感知、边缘智能、农场决策引擎、执行、农户咨询),边缘网关运行量化EfficientNet、MLP与土壤湿度预测器综述引用的既有研究,包括实验室叶片图像、土壤检测数据与灌溉试验;原文未提供本综述自建的实验数据判断病害CNN与土壤湿度预测成熟度高,养分模型依赖传感器精度,强化学习灌溉为新兴方向;结论为集成AI系统可增产并节水、节肥、减药综述性质,未开展原始实验,未报告任何准确率、延迟或节水实测值;指出实验室到田间迁移差、低成本养分传感器精度不足、农村网络连通性差等挑战
AI-Enabled Sustainable Precision Farming Using Raspberry Pi and Coral Edge TPU with Cloud Based Model Training

[2][2]

树莓派+Coral Edge TPU的作物状态分类与IoT传感系统(仅摘要)云端训练深度学习模型后量化并转换为边缘优化格式,部署于Coral Edge TPU;采集作物照片与土壤湿度、温度、湿度、光照数据摘要未提供数据集规模、作物种类与测试集划分细节报告准确率94.82%、精确率94.47%、召回率94.05%、F1 94.26%;推理延迟约38毫秒、吞吐26.3张/秒;据示例性可持续性研究估计可节省每日灌溉用水22%,该22%为示例性估计而非实测仅摘要可得;节水22%为示例性估计;未报告田间部署验证;测试集规模与作物构成缺失,结果仅在该摘要所述条件下成立
IoT-Based Weather Monitoring Station for Precision Agriculture Using Arduino Uno

[3]

基于Arduino Uno的IoT气象监测站(仅摘要,无量化性能指标)Arduino Uno配合DHT11、BMP180、土壤湿度、雨滴传感器与LCD显示,可选ESP8266 Wi-Fi模块上传IoT平台摘要未提供实验规模、部署地点与对照条件称传感器读数准确、数据处理可靠、环境监测高效,系统成本低、易安装、节能,适合中小规模应用;未报告精度、延迟或节水数值仅摘要可得;无边缘AI推理能力;不提供可量化的部署性能数据,不能与其他来源的准确率或延迟比较
Edge AI for Autonomous Agricultural Drone Data Processing: Onboard Real-Time MobileNetV3 Inference with RPLiDAR Sensor Fusion for Instant In-Flight Crop Health Monitoring

[4][4]

无人机机载MobileNetV3推理与RPLiDAR传感器融合的作物健康监测框架(全文,端到端仅仿真)MobileNetV3-Large微调后机载推理,RPLiDAR估计株高与冠层体积,融合为0–100健康指数;GPS自主覆盖、Flask后端、Three.js仪表板与自动PDF报告8400张卡纳塔克邦三农场椰子棕榈真实航拍图像微调,1260张留出真实影像测试;端到端流程在AirSim仿真中五次任务验证三类准确率98.2%、92.5%、95.8%;LiDAR株高/冠层体积MAE 14.3厘米(对模拟真值);仿真中平均18.4分钟覆盖全田,Flask API响应22毫秒,PDF生成3.1秒;62 FPS为RTX 3060代理基准端到端验证仅在仿真中完成;物理Jetson NX基准、热特性与田间验证未完成;LiDAR误差对照为模拟真值;分类任务为椰子冠层三分类,与树莓派研究的作物状态分类任务不同,数值不可直接比较

关联文章

4 篇

核心材料

Arduino物联网气象站支撑田间实时监测

论文Zenodo (CERN European Organization for Nuclear Research)一手来源

IoT-Based Weather Monitoring Station for Precision Agriculture Using Arduino Uno

基于Arduino Uno搭建低成本物联网气象站,实时监测农田环境参数。

关注点

该方案用DHT11、BMP180、土壤湿度与雨量传感器配合Arduino Uno实现田间环境实时监测,硬件成本低、易部署,但传感器精度有限,仅适合小中型农田的入门级应用。

阅读 0

树莓派与边缘TPU精准农业系统,准确率94.82%

论文Figshare一手来源

AI-Enabled Sustainable Precision Farming Using Raspberry Pi and Coral Edge TPU with Cloud Based Model Training

构建基于树莓派与Coral Edge TPU的边缘AI精准农业系统,云端训练模型并本地实时决策。

关注点

该方案在Coral Edge TPU上实现约38ms推理延迟与26.3张/秒吞吐,并估算按需灌溉可节水22%,可作为边缘AI农业部署的性能参照,但节水数据来自估算而非田间实测。

阅读 0

背景阅读

边缘病害诊断与精准灌溉综述,提供背景

论文Zenodo (CERN European Organization for Nuclear Research)

Ai-Based Smart Agriculture: A Review Of Edge Disease Diagnosis, Soil Nutrient Profiling And Reinforcement Learning For Precision Irrigation

Ai-Based Smart Agriculture: A Review Of Edge Disease Diagnosis, Soil Nutrient Profiling And Reinforcement Learning For Precision Irrigation。Zenodo (CERN European Organization for Nuclear Research)

阅读 0

深入理解

主体、方法与延伸问题
理论方法

量化与轻量骨干网络:各自未验证之处

延伸思考

综述提出边缘网关运行量化EfficientNet以支持离线诊断,但未提供量化后的精度损失数据;无人机研究选用MobileNetV3-Large作为机载轻量骨干,其62 FPS吞吐来自RTX 3060桌面GPU代理而非目标Jetson NX。两者无共同实验条件,不能合并为同一“轻量化—精度—延迟”权衡结论;各自模型在目标硬件上的精度与延迟仍需分别实测。[1][4]

延伸问题

边缘方案的田间精度与节水效果尚缺实测

现有来源中,树莓派+Edge TPU系统的94.82%准确率与22%节水估计均出自摘要,节水被明确称为示例性研究估计;无人机框架的端到端性能来自仿真;Arduino气象站未报告量化指标。因此,这些边缘架构在真实田间的分类精度、端到端延迟与灌溉节水效果是否成立,仍是未解问题。[2][4]

相关主体

传感融合:从单一图像分类到结构指标融合

延伸思考

无人机研究将相机分类置信度与LiDAR株高、冠层体积融合为单一0–100健康指数,代表从纯图像分类向多模态融合的延伸;但LiDAR误差是对模拟真值评估的,融合指数的有效性尚未在真实田间确认,其增益不能外推到其他作物或平台。[4]

参考资料

每条发现都可以回到来源
  1. 1
    Ai-Based Smart Agriculture: A Review Of Edge Disease Diagnosis, Soil Nutrient Profiling And Reinforcement Learning For Precision Irrigation

    Zenodo (CERN European Organization for Nuclear Research) · 站内文章 · 依据已获取正文片段

    查看引用摘录
    Disease CNNs and soil-moisture prediction are mature, nutrient models depend on sensor accuracy, and RL irrigation is promising but needs careful validation.
    A gateway runs a quantized EfficientNet for disease diagnosis, an MLP for nutrient classification and a soil-moisture forecaster, working offline when needed.
    G2: Models trained on laboratory images and foreign soils transfer poorly to Indian fields.
  2. 2
    AI-Enabled Sustainable Precision Farming Using Raspberry Pi and Coral Edge TPU with Cloud Based Model Training

    Figshare · 站内文章 · 依据摘要

    查看引用摘录
    The optimized model achieves an exemplary accuracy of 94.82%, precision of 94.47%, recall of 94.05% and F1-score of 94.26%.
    Deploying with the Coral Edge TPU reduces inference latency to approximately 38ms compared to Raspberry Pi-only processing, increases processing throughput to 26.3 pictures per second and reduces CPU usage.
    Illustrative sustainability study estimates that condition based monitoring can save 22% of the irrigation water used every day.
  3. 3
    IoT-Based Weather Monitoring Station for Precision Agriculture Using Arduino Uno

    Zenodo (CERN European Organization for Nuclear Research) · 站内文章 · 依据摘要

    查看引用摘录
    Experimental testing demonstrated accurate sensor readings, reliable data processing, and efficient environmental monitoring.
    The proposed system is affordable, easy to install, energy-efficient, and suitable for small- and medium-scale agricultural applications
  4. 4
    Edge AI for Autonomous Agricultural Drone Data Processing: Onboard Real-Time MobileNetV3 Inference with RPLiDAR Sensor Fusion for Instant In-Flight Crop Health Monitoring

    Zenodo (CERN European Organization for Nuclear Research) · 站内文章 · 依据已获取正文片段

    查看引用摘录
    physical Jetson NX benchmarking, thermal characterisation, and field validation remain to be completed
    A MobileNetV3-Large network, fine-tuned on 8,400 real aerial images of coconut palms from three farms in Karnataka and evaluated on a 1,260-image held-out real-imagery test set, classifies crop canopies as Healthy, Stressed, or Diseased with per-class accuracies of 98.2%, 92.5%, and 95.8% respectively.
    an RPLiDAR sensor estimates plant height and canopy volume with a mean absolute error of 14.3 cm against simulated ground truth
    was validated end-to-end across five missions in the Microsoft AirSim simulation environment, completing full-field coverage in an average of 18.4 minutes
    Inference throughput of 62 FPS was benchmarked using the TensorRT FP16 engine on the simulation workstation's RTX 3060 GPU as a proxy for the target NVIDIA Jetson NX platform; physical Jetson NX benchmarking, thermal characterisation, and field validation remain to be completed.
    classifies crop canopies as Healthy, Stressed, or Diseased with per-class accuracies of 98.2%, 92.5%, and 95.8% respectively
    Inference throughput of 62 FPS was benchmarked using the TensorRT FP16 engine on the simulation workstation's RTX 3060 GPU as a proxy for the target NVIDIA Jetson NX platform
    it is presented here as a simulation-validated proof-of-concept pending real-hardware and real-field confirmation
    these structural metrics are fused with the camera-derived classification confidence into a single 0-100 health index

更新记录

  • 建立专题,整理关联文章与研究背景

本专题由 AI 基于引用资料综合整理;“延伸思考”为分析性推论。研究结果的适用范围与引用请以原文为准。