Aplicación de Inteligencia Artificial sobre infraestructuras IoT para automatizar y optimizar los procesos de agricultura intensiva en invernaderos

  1. Juan Morales García
Supervised by:
  1. Andrés Bueno Crespo Director
  2. José María Cecilia Canales Director

Defence university: Universidad Católica San Antonio de Murcia

Fecha de defensa: 05 June 2023

  1. Andrés Muñoz Ortega Chair
  2. Antonio Llanes Castro Secretary
  3. José Santa Lozano Committee member

Type: Thesis

Teseo: 814693 DIALNET lock_openTESEO editor


The United Nations Sustainable Development Goals (SDGs) establish a series of goals aimed at eradicating poverty, protecting the planet and ensuring prosperity for its citizens. These goals include: (6) “To ensure availability and sustainable management of water and sanitation for all”, (13) “To take urgent action to combat climate change and its impacts” and (15) “To protect, restore and promote sustainable use of terrestrial ecosystems, sustainably manage forests, combat desertification, halt and reverse land degradation and halt biodiversity loss”. Industrial processes and, in particular, intensive agricultural processes, are one of the main threats to meeting the SDGs. However, technological advances in areas such as Artificial Intelligence (AI), High Performance Computing (HPC) or the Internet of Things (IoT) allow increasing the productivity of these processes reducing their environmental and ecosystemic impact. The research developed in this doctoral thesis aims to establish a framework where to take advantage of the technological advances developed in these disciplines, i.e. AI, HPC and IoT, to optimize and reduce the impact of the most environmentally aggressive industrial processes. Specifically, this PhD thesis will be developed in the context of intensive greenhouse agriculture, a sector of great strategic, commercial and even humanitarian value to ensure access to food for all humanity, focusing on three key points: (1) the generation of low-power AI techniques that can be executed on platforms with reduced computational capabilities, such as IoT devices; (2) the creation of an infrastructure that allows training, deploying and predicting with AI techniques that require large computational capabilities on small IoT devices thanks to real-time communication protocols such as MQTT; and (3) the increase of computational capabilities and energy efficiency of IoT devices thanks to the virtualization of remote GPUs through rCUDA. The main results obtained in relation to the above demonstrate that (1) the intersection between AI, HPC and IoT is still very nascent. The computational loads of machine learning are becoming higher and higher and increasingly diverge from the computational resources available on the computing devices closest to the data capture, i.e., edge computing devices. These platforms are not computationally capable of performing some of the most demanding tasks (such as, for example, training AI techniques), limiting the success of their application; (2) an auxiliary infrastructure can be created to develop real-time predictions in IoT devices, although the exchange of information between the different nodes of the infrastructure involves an assumable latency since it is very low; and (3) it is possible to extend the computational capabilities and energy efficiency of IoT devices through the use of remote GPU virtualization techniques. These techniques significantly increase the energy efficiency of these devices by delegating the most computationally intensive operations to remote compute servers. Although it is true, the total consumption of the infrastructure increases significantly due to the communication costs between the devices it edge and it cloud. Finally, it should be noted that this thesis has been developed in the challenge-collaboration project “Development of high performance IoT infrastructures against climate change based on artificial intelligence” (GLOBALoT) with reference RTC2019-007159-5, funded by the Ministry of Science and Innovation / State Research Agency, which has a strong technological character and, therefore, the knowledge obtained has been transferred, developing a functional prototype in TRL 3-4 that has been deployed in a real greenhouse environment offered by one of the project partners, the company NUTRICONTROL. The results obtained show a clear interest in this technology, laying the foundations to automate and optimize processes through Artificial Intelligence of Things (AIoT) to increase production and reduce environmental impact in smart greenhouses.