Tiny machine learning on the edge: A framework for transfer learning empowered unmanned aerial vehicle assisted smart farming
Tiny machine learning on the edge: A framework for transfer learning empowered unmanned aerial vehicle assisted smart farming
复制标题
DOI：
10.1049/smc2.12072
复制
发表时间：
2023-11
期刊：
影响因子：
3.1
通讯作者：
A. Hayajneh;Sami A. Aldalahmeh;Feras Alasali;H. Al-Obiedollah;Sayed Ali Zaidi;Des McLernon
中科院分区：
文献类型：
--
作者：
A. Hayajneh;Sami A. Aldalahmeh;Feras Alasali;H. Al-Obiedollah;Sayed Ali Zaidi;Des McLernon
新兴技术正在不断重新定义智能农业的范式，并为更精确和更明智的农业实践开辟道路。提出了一种基于微型机器学习（TinyML）的框架，用于无人机（UAV）辅助的智能农业应用。演示了在无人机和定制物联网（IoT）传感器上实际部署这种框架，这些传感器测量土壤湿度和周围环境条件。该框架的主要目标是利用TinyML使用深度神经网络（DNN）和长短期记忆（LSTM）ML模型实现迁移学习（TL）。作为一个案例研究，该框架用于预测智能农业应用的土壤含水量，通过时间序列预测模型指导作物的最佳用水。据作者所知，以前还没有研究过使用TinyML将无人机辅助TL用于边缘物联网的框架。基于TL的框架在不同但相似的应用程序和数据域上使用预训练的数据模型。作者不仅展示了所提出的框架的实际部署，而且还通过真实的部署量化了其性能。这是通过设计用于土壤和环境传感的定制传感器板来实现的，该传感器板使用ESP32微控制器单元。针对边缘设备上的不同ML模型架构以及其他性能指标（即均方误差和决定系数[R2]）测量了推理指标（即推理时间和准确性），同时强调了平衡准确性和处理复杂性的需求。总之，结果显示了使用无人机将DNN和LSTM模型的TL交付给超低性能边缘物联网设备进行土壤湿度预测的实际可行性。但总的来说，这项工作也为进一步研究TinyML在智能农业许多不同方面的其他应用奠定了基础。
Emerging technologies are continually redefining the paradigms of smart farming and opening up avenues for more precise and informed farming practices. A tiny machine learning (TinyML)‐based framework is proposed for unmanned aerial vehicle (UAV)‐assisted smart farming applications. The practical deployment of such a framework on the UAV and bespoke internet of things (IoT) sensors which measure soil moisture and ambient environmental conditions is demonstrated. The key objective of this framework is to harness TinyML for implementing transfer learning (TL) using deep neural networks (DNNs) and long short‐term memory (LSTM) ML models. As a case study, this framework is employed to predict soil moisture content for smart agriculture applications, guiding optimal water utilisation for crops through time‐series forecasting models. To the best of authors’ knowledge, a framework which leverages UAV‐assisted TL for the edge internet of things using TinyML has not been investigated previously. The TL‐based framework employs a pre‐trained data model on different but similar applications and data domains. Not only do the authors demonstrate the practical deployment of the proposed framework but they also quantify its performance through real‐world deployment. This is accomplished by designing a custom sensor board for soil and environmental sensing which uses an ESP32 microcontroller unit. The inference metrics (i.e. inference time and accuracy) are measured for different ML model architectures on edge devices as well as other performance metrics (i.e. mean square error and coefficient of determination [R2]), while emphasising the need for balancing accuracy and processing complexity. In summary, the results show the practical feasibility of using drones to deliver TL for DNN and LSTM models to ultra‐low performance edge IoT devices for soil humidity prediction. But in general, this work also lays the foundation for further research into other applications of TinyML usage in many different aspects of smart farming.