FRESQIA: Foliar symptom recognition with quantum artificial intelligence

Quantum innovation for precision agriculture with FRESQIA

Foliar diseases pose a major threat to productivity and phytosanitary management, especially because of their rapid spread. Manual inspection is limited because of its subjective nature and because it is a slow and difficult process to scale up to large crop areas. Therefore, when symptoms are clearly visible, it may be too late to apply effective control measures.

To overcome these limitations, the proposed solution implements automatic monitoring using machine vision systems, enabling agronomic monitoring of the crop. This approach integrates quantum computing techniques for the analysis of complex visual patterns, contrasting classical models with Quantum Convolutional Neural Networks (QCNN).

The goal is an early detection of symptoms, with potential to reduce pesticide use and increase agricultural yield. In this work, the case study and images used correspond to strawberry leaves in greenhouses, although the proposed approach is applicable to other crops with visible leaf symptomatology.

Preliminary results show good performance with favorable illumination and clearly visible symptoms, and limitations in cases of leaf overlap and partially hidden leaves. The paper discusses these limitations and highlights the potential of compact quantum models to support agronomic monitoring.

Highlighted Results

Within the framework of this demonstrator, we have validated our technology through the following practical workflow:

• Early detection and leaf isolation.

  We use RGB cameras in controlled environments (relying on CTIC's Climate Simulator) and apply classical object detection models to automatically identify and trim leaves. This allows us to handle complex challenges of the real environment, such as leaf overlap or background elements, ensuring that the system focuses only on relevant patterns.

• Symptom classification with QCNN.

  We implement Quantum Convolutional Neural Networks (QCNN) to analyze images and classify whether leaves are diseased. By comparing this quantum model with classical convolutional networks (CNNs), we demonstrate that the quantum approach achieves equivalent performance using only a tiny fraction of parameters (just a few hundred) and considerably fewer training images.

• Compact models for agriculture.

  We validate the technical feasibility of hybrid-quantum solutions in the agri-food sector. Our architecture is very compact, requires fewer training images and consumes significantly less memory than classical models. This structural efficiency facilitates its future direct integration into field devices and ensures high scalability as quantum hardware evolves.

Conclusions

In the agricultural domain, the system can contribute to reducing crop losses due to foliar diseases and pesticide dependency, as well as eliminating the bottleneck of manual inspection through objective and standardized surveillance.

Technologically, the feasibility of hybrid-quantum solutions is validated, thanks to a very compact architecture that requires fewer training images and consumes significantly less memory than classical models. This efficiency facilitates their direct integration into agricultural devices and ensures high scalability, with a potential for continuous improvement as quantum hardware evolves.