Hardware

Motivation

One of the main challenges in tackling Greening is that Candidatus Liberibacter asiaticus, the bacterium responsible for the disease, cannot be cultivated under laboratory conditions. Traditional diagnostic methods, such as real-time PCR, are accurate but slow, expensive, and require specialized laboratory infrastructure. This creates a significant challenge for farmers and researchers trying to protect orchards and develop effective treatments.

The need for a rapid diagnostic tool inspired the creation of the Sniffer. We wanted a device capable of quickly and accurately detecting Greening in a lab, and even maybe directly in the orchard. During our research, we found that dogs can detect Greening with extraordinary precision. A study evaluating ten canines trained to identify Candidatus Liberibacter asiaticus (CLas) reported 0.9905 accuracy, 0.8579 sensitivity, and 0.9961 specificity^[1]. Remarkably, these dogs could detect cryptic infections within two weeks post-infection, while visual inspection or qPCR could take 1 to 32 months to identify the same cases. This curious phenomenon motivated a key question: Could we create a system that mimics a dog’s sense of smell? The Sniffer was born from this idea.

VOCs

To translate the biological insight into a practical sensing strategy, we focused on volatile organic compounds (VOCs) emitted by citrus plants. VOCs are carbon-based molecules that readily evaporate at room temperature and are continuously released as part of normal plant metabolism. Their composition and abundance vary with species, developmental stage, and physiological condition. Importantly, studies have shown that VOC emission profiles change significantly when plants experience stress, including pathogen infections. In the case of Greening, infected citrus trees produce a distinct VOC signature, likely reflecting alterations in systemic defense responses^[2].

Based on this knowledge, our hardware was specifically designed to detect and analyze VOC emission patterns, allowing us to distinguish between healthy and infected plants. By capturing these subtle chemical differences, the Sniffer can replicate the dogs’ ability to identify infected trees, but in an electronic and lab-safe format.

Nanotech

To achieve the sensitivity and selectivity required, we turned to nanotechnology, which has proven to be transformative in biology and disease detection. At the nanoscale, materials exhibit unique physical and chemical properties that do not appear in bulk, allowing for highly specific interactions with biological molecules and chemical compounds. This makes nanostructured surfaces ideal for detecting subtle chemical signatures, such as the VOCs emitted by citrus leaves infected with CLas.

In the Sniffer, this approach is realized through a custom polymer-based sensor composed of multiple interdigitated electrodes. These electrodes act like tiny capacitors: when VOCs from citrus leaves interact with the electrodes, the local electric field changes, producing a measurable electrical signal.

Three of the IDEs are coated with nanostructured polymer films assembled by the Layer-by-Layer (LbL) technique, while one remained uncoated as reference. This arrangement creates multiple complementary sensing units, allowing the device to generate a distinct electrical “fingerprint” for each sample, much like a dog’s nose integrates multiple chemical cues.

Electronic Circuit

The Sniffer’s polymer-based sensors are only part of the story—their responses need to be measured, processed, and converted into meaningful information. This is the role of the electronic circuit, designed to generate excitation signals, read the sensors, and digitize the responses for analysis. For the polymer-based IDEs, the circuit applies a sinusoidal AC signal, allowing the device to capture both capacitive and resistive changes induced by VOC interactions. The local changes in the electric field of the electrodes, caused by VOC adsorption, are transformed into measurable electrical signals in real time. At the heart of the system is the ESP32 microcontroller, which coordinates signal generation, collects digital data, and prepares it for further processing using machine learning. Additional components, such as amplifiers, multiplexers, and high-resolution analog-to-digital converters, ensure that even subtle variations in sensor response are accurately captured.

Complementary Sensors

To enhance the Sniffer’s detection capabilities, we incorporated several complementary sensors alongside the nanofabricated polymer-based units:

• MQ series (MQ-3, MQ-135, MQ-136, MQ-137): Metal-oxide semiconductor (MOS) gas sensors that respond to specific groups of gases by changing their electrical resistance. They provide additional information about volatile compounds, complementing the polymer sensor’s VOC fingerprint.
• BME680: An environmental sensor that measures temperature, humidity, pressure, and total VOCs. Including BME680 data helps account for environmental variations that may influence VOC emissions or sensor responses.
• SHT31: A high-precision sensor for temperature and relative humidity, providing additional environmental context to refine VOC measurements.

By combining these complementary sensors with the polymer-based IDEs, the Sniffer can integrate chemical and environmental data, improving the accuracy and reliability of disease detection in citrus trees.

Air pumping

Of course, to efficiently deliver VOCs from citrus leaves to the sensor array, we had to incorporate an air pumping system. A miniature DC pump drives a stream of air through silicone tubing and a custom PDMS microchannel, ensuring even distribution of VOCs across the sensors. Humidity filter valves at the inlet and outlet protect the system from moisture, preventing condensation and preserving sensor stability. Finally, we designed a 3D-printed case to house the entire system, with two compartments that separate electronic circuits, sensors, and airflow pathways from the leaf chamber.

Bill of Materials

Firmware

The firmware is the brain of the Sniffer, a specialized software that runs directly on the ESP32 microcontroller and manages all hardware operations.. To avoid any interference, the firmware operates with two main tasks in parallel: one dedicated exclusively to reading all sensors and performing the frequency sweep, and another responsible for sending the complete data packets via Bluetooth Low Energy (BLE) to a computer or smartphone.

Additionally, a specific firmware was developed for the calibration of the MQ-series gas sensors. This process, which lasts several hours, establishes the baseline resistance of each sensor in clean air (R0). This value is permanently saved to the device's memory and serves as a reference to normalize future readings, ensuring that measurements are consistent and comparable over time.

Machine Learning

While the Sniffer hardware can capture subtle chemical variations in citrus VOC emissions, raw sensor readings alone are not sufficient for reliable disease detection. The challenge lies in translating electrical signals into actionable insights — a task perfectly suited for machine learning. The process follows some steps:

Data Collection and Preparation

To build a dataset, we established a standardized methodology for sample collection. Fresh orange leaves were collected and categorized as either Greening-positive or Greening-negative. Each leaf was sliced to maximize VOC emission and placed inside the sterilized Sniffer chamber. The miniature air pump delivered VOCs evenly across the sensor array while the ESP32 recorded sensor signals.

Each measurement lasted approximately 25 minutes to ensure signal stabilization and capture a complete chemical profile. The collected data, including readings from nanostructured polymer sensors, MOS gas sensors, and environmental sensors, were stored as CSV files for further processing.

Before feeding the data into machine learning models, a preprocessing pipeline cleaned and structured the dataset. Duplicate rows were removed, missing values interpolated, and extreme outliers clipped to reduce noise. Data balancing techniques ensured that classes were proportionally represented, preventing model bias toward the majority class. Finally, the dataset was split into training, validation, and test sets, either randomly at the row level or grouped by original CSV file to evaluate generalization across different measurement sessions.

Model Training

A variety of classification algorithms were trained to discriminate between Greening-positive, Greening-negative, and empty (baseline) samples. These included Logistic Regression, Support Vector Machine (SVM), Random Forest, XGBoost, PCA combined with KMeans, and Linear Discriminant Analysis (LDA). Each of these models "learns" to associate a specific pattern of sensor responses with a diagnosis.

Training the Model

During training, each model was fitted on the training set, validated against the validation set, and tested with unseen data. Performance metrics calculated included accuracy, balanced accuracy, precision, recall, F1-score, and ROC-AUC. Visual outputs such as confusion matrices and ROC curves were automatically generated for interpretability. After training, the models are rigorously evaluated with data they have never seen before to verify their accuracy and reliability. The best-performing model is selected and saved.

Diagnosis (Inference)

With a trained model in place, the Sniffer can perform real-time diagnosis. New sensor readings are automatically preprocessed, aligned with the model’s expected input, and classified as positive, negative, or empty. When a new measurement is taken, the data is sent to a computer, where the machine learning model analyzes it and provides a diagnosis in seconds, reporting whether the sample is positive or negative for the disease.

Results

After testing multiple combinations of models and configurations, the most promising approach was found to be Linear Discriminant Analysis (LDA). LDA serves both as a classification method and a dimensionality reduction technique. It projects the data into a lower-dimensional space, maximizing the separation between classes while minimizing the spread within each class. The sensor data showed significant variance, making it suitable for building a discriminative classification model.

The training results was encouraging: the model can successfully distinguish between healthy and infected leaves. However, the system remains sensitive to environmental variations. This is mainly due to the limited size of the dataset and the lack of generalization across different environmental conditions. Measurements taken on different days, for instance, can affect the model’s performance.

Importantly, this sensitivity is not necessarily a result of overfitting. The models show strong performance on the test sets, confirming that the dataset split into training, validation, and test subsets provides a reliable evaluation. The confusion matrix offers a clear visual representation of the model’s ability to classify the leaves correctly and highlights areas where misclassifications occur.

A confusion matrix helps you evaluate a classification model's performance. The rows typically represent the actual (true) labels, while the columns show the predicted labels. The main diagonal (from top-left to bottom-right) contains the correctly classified instances, meaning the predicted label is the same as the actual label. All other cells off the diagonal represent errors, showing where the model confused one class for another (misclassifications).

Table I. Sniffer precision and recall

This table is a classification report that summarizes the excellent performance of a predictive model for the "negative" and "positive" classes. The Precision scores (0.97 and 0.98) indicate that when the model predicts a class, it is highly accurate, leading to a very low false positive rate. The Recall scores (0.98 and 0.97) show that the model is extremely effective at identifying the vast majority of actual instances for each class, meaning it has a low false negative rate. Consequently, the F1-Score, which is the harmonic mean of precision and recall, is an outstanding 0.98 for both classes, confirming a robust and well-balanced performance. The Support column simply indicates that this evaluation was based on 121 actual instances of the 'negative' class and 131 of the 'positive' class.

Final Prototype Overview

After months of iteration, sketches, and conversations with experts, the different pieces of our design finally came together into a single device: the Sniffer prototype. At its core, the system follows a simple flow. Citrus leaves are placed inside a chamber, where a miniature pump gently pulls air across them. This airflow carries the volatile compounds released by the leaves into a microchannel that distributes them evenly across the sensor array.

Inside the sensing chamber, two worlds meet. On one side, our nanostructured polymer sensors capture subtle chemical variations in the air, translating them into tiny changes in capacitance. On the other, commercial gas sensors complement these measurements by responding to broader families of gases. Both signals are collected by the electronic circuit, digitized, and sent to the ESP32, the microcontroller that acts as the brain of the system.

The entire manufacturing process and the physical principles behind the hardware's operation are thoroughly documented in our Handbook and in our Photo Gallery.

Future Improvements

While the Sniffer prototype already demonstrates its ability to discriminate between healthy and infected citrus leaves, we recognize that there are many opportunities to make the device even more robust, portable, and versatile. Some of the improvements we envision include:

• Noise Isolation with a Metal Enclosure

To reduce external electromagnetic interference and improve signal stability, future versions could be housed in a shielded metal case.

• Expanding Training Datasets

The machine learning model becomes more reliable as it is trained with larger and more diverse datasets. Collecting thousands of samples across different seasons, orchards, and tree ages will significantly improve predictive accuracy.

• Improving Model Robustness
We also have specific suggestions to develop a better machine learning model:
• Use more comprehensive datasets to achieve better generalization across different environments and a wider range of leaves.
• Use whole leaves for sampling, avoiding the indirect “overfitting” effect caused by using multiple slices from the same leaf, which can falsely inflate data diversity.
• Reduce the time between leaf collection and measurement, ensuring that VOCs are captured as close as possible to the moment of extraction for more accurate and reliable data.
• Testing with a Wider Range of Diseases

While the prototype focuses on Citrus Greening, additional tests with other citrus pathogens — and even unrelated plant diseases — will help validate the flexibility of the system.

• Battery Integration for Field Deployment

A built-in rechargeable battery would make the Sniffer fully portable, allowing direct, real-time diagnosis in the field without dependence on external power.

• Optimized Airflow Control

Integrating sensors and actuators to fine-tune the air pumping system could ensure that VOC delivery to the sensing chamber is consistent, reproducible, and optimal for detection.

• Miniaturization of Electronics

Future iterations could reduce circuit footprint and integrate more components into a single PCB, improving portability and usability for farmers.

• User-Friendly Software for Direct Diagnosis

A key step forward will be the development of an intuitive software interface. Instead of raw data and graphs, users would receive a clear, straightforward diagnostic output (e.g., “Healthy” or “Infected”), making the tool accessible to farmers, technicians, and policymakers without specialized training.

Development Journey

The creation of the Sniffer was a journey defined by collaboration, persistence, and continuous learning. From the very beginning, our team spent countless nights in the lab— often until sunrise — testing circuits, reassembling boards, and debugging issues. Every step of the process required resourcefulness and resilience. Because many of the components and PCBs needed for our system are not readily available in Brazil, we had to import almost everything, facing extremely high import taxes and long waiting times due to national trade policies.

Despite these obstacles, we remained committed to building a precise and reliable system. Working with SMD components was especially challenging — soldering under a microscope, dealing with fragile connections, and ensuring minimal noise and interference demanded patience and precision. However, the use of SMD was absolutely essential to guarantee the accuracy, compactness, and stability that the project required.

Along the way, some components failed unexpectedly, forcing us to redesign parts of the system or wait weeks for replacements. These setbacks, while frustrating, became key learning experiences that strengthened our understanding of both the electronics and the physics underlying our measurements. The diagram below highlights the steps we followed, from the earliest brainstorms to the final evaluation of our objectives.

This project would not have been possible without the generosity and support of our collaborators, who kindly donated imported components and shared their technical expertise when we needed it most. Their contributions — and the collective effort of everyone involved — transformed what once seemed like an impossible challenge into a functioning, data-driven prototype capable of addressing real agricultural problems.

Ultimately, building the Sniffer was much more than a technical achievement — it was a story of perseverance in the face of limited resources, proof that innovation can thrive even under constraints when driven by passion, teamwork, and shared purpose. The diagram below highlights the steps we followed, from the earliest brainstorms to the final evaluation of our objectives.

Universal detection tool

The results of our machine learning experiments confirmed that the Sniffer is not only functional but also highly effective. Measurements from the same leaf were consistent, and our models were able to reliably distinguish between healthy and infected leaves, demonstrating that the hardware and software work together as a cohesive system. These outcomes serve as a proof of concept, showing that the Sniffer could be a valuable tool for farmers and researchers alike, providing fast and accurate diagnostics directly in the field.

One of the greatest strengths of the Sniffer lies in its flexibility. The entire sensing circuit is modular, allowing each sensor unit to be replaced or reconfigured without redesigning the device. Firmware can be easily adapted to new modules, and the machine learning algorithms can be retrained on fresh datasets, enabling the system to learn and discriminate new patterns with high precision. This makes the Sniffer a universal detection tool - a platform that can be tailored to monitor a wide variety of VOC-based signals beyond Citrus Greening. Whether for other agricultural pathogens, environmental monitoring, or biomedical applications, the Sniffer provides a customizable framework where hardware, software, and algorithms can be adapted to meet new challenges.

By combining a successful proof-of-concept with modular design and adaptable machine learning, our work demonstrates that innovation thriving when precision, flexibility, and practical utility come together. The Sniffer is not just a prototype - it is a innovation thrives when precision, flexibility, and practical utility come together, ready to tackle a wide range of real-world detection challenges.

Handbook

The development of the injection hardware was driven by the necessity to devise a novel and precise technique for delivering peptides into citrus plants afflicted with Greening disease. In contrast to conventional methods such as spraying or soil application, the present solution facilitates direct injection into the plant's vascular system. This strategy is intended to ensure greater bioavailability of the peptide, thereby optimising its action and reducing losses during application.

Furthermore, the system's design was developed to maximize peptide utilisation, thereby preventing leaks and minimizing the dispersion of material into the environment. This feature enhances therapeutic efficacy and contributes to a more sustainable and safer application, reducing potential environmental impacts. Consequently, the hardware constitutes a viable and scalable alternative, aligned with contemporary agricultural practices and the pressing need for innovative solutions to combat Greening disease.

Needles

Description

The selection of the needles was the result of a comprehensive review of the scientific literature and other technical sources, with the objective of identifying the most efficient and minimally invasive strategy for peptide delivery in citrus plants. Needles offer significant advantages over traditional methods, as they can penetrate the plant epidermis in a controlled manner and deliver the solution directly into the vascular tissue, ensuring greater efficiency in transporting the target molecule to its site of action.

The design was conceived to optimize liquid flow and application precision, adopting open-channel needles that facilitate continuous flow of the injected solution. The structural parameters, such as dimensions, insertion angle, and channel density, were defined in accordance with a comprehensive literature review and theoretical modelling. These definitions were then complemented by practical estimations, with the objective of meeting the specific needs of the project.

3D Printing

In the development of the prototype using additive manufacturing, Fusion 360 software was utilised for the detailed modelling of the needles and microchannel system, which facilitates the connection between the reservoir and the application point. This software enhanced precision in the control of dimensions, angles, and structural parameters, which are pivotal in ensuring the efficacy of peptide flow during injection.

The files generated from these models are available for 3D printing and can be modified by other teams or researchers interested in improving or adapting the system to their own needs, thereby strengthening the open and collaborative nature of the initiative.

Reservatory

Description

The injection equipment reservoir was designed with the primary objective of providing portability and safety to the system, thereby enabling the storage and transport of the peptide that will be used in plant treatment. The reservoir has been designed to be compact and lightweight, thus enabling the hardware to be moved between different areas of the field or nursery without compromising the integrity of the contents. The geometry and attachment of the device ensure balance during transport and operation, thereby reducing vibrations and minimising the risk of spillage or contamination.

This concern with containment prevents material loss during transport and application, ensuring greater efficiency in the process and reducing waste that could compromise both the cost and sustainability of the operation.

3D Printing

In the development of the reservoir prototype through additive manufacturing, Fusion 360 software was employed for detailed modelling of its geometry and structural features. This platform enabled precise control over dimensions, wall thickness, and attachment points, ensuring that the reservoir could achieve the intended objectives of portability and safety while maintaining compatibility with the rest of the injection hardware. The CAD environment also facilitated the incorporation of smooth transitions between the square storage chamber and the cylindrical base, ensuring a balanced structure that can be easily integrated into the application system.

For 3D printing, the design was exported in STL format and prepared using slicing software, where parameters such as layer height, infill density, and wall thickness were adjusted to balance mechanical resistance with lightweight construction. The printing process can be carried out using PLA or PETG filaments, in our case, we used ABS, these are materials that provide sufficient rigidity and durability, while also being compatible with biotechnological prototyping environments.

This workflow allows the reservoir to be reproduced and adapted by other teams or researchers, supporting the collaborative improvement of the system and ensuring its adaptability for diverse experimental conditions in plant peptide treatments.

Tests and Results

In order to validate the performance of the delivery system, tests were conducted. The experiment employed plant branches as a model. The initial step in the experimental process was to assess the needles resistance and penetration power, in order to verify whether they would be capable of piercing the stem efficiently. The findings of the tests demonstrated that the needles exhibited sufficient strength and precision to penetrate the plant tissue, thereby ensuring access to the plant's vascular system.

Subsequently, a solution of water mixed with dye was prepared and utilised as a marker to track the internal transport of the liquid post-injection. The solution was administered via the delivery system, and after a brief interval, incisions and cross-sections were made in the branch to observe the distribution of the dye. Observations confirmed the presence of the liquid inside the conducting vessels, thereby demonstrating that the injected fluid was effectively transported by the plant's vascular system.

Consequently, it was determined that the tests were successful the needles were able to penetrate the stem without compromising its structural integrity, and the injected liquid reached the vascular system, validating the operating principle of the injection prototype.

Improvements to be made

With regard to future improvements to the device, the main opportunity for advancement relates to the choice of manufacturing material. In the initial prototyping stages, printing on common polymers was sufficient; however, greater robustness and durability of the device could be ensured through the use of more resistant materials, such as metals, in subsequent production stages. This modification would serve to reduce the likelihood of deformation or breakage during the injection process. Furthermore, it would expand the range of application sites, for instance enabling direct insertion into the trunk of plants without compromising the structural integrity of the system.

Another potential avenue for enhancement involves the utilisation of higher precision and scale 3D printers, which facilitate the fabrication of components in diverse dimensions, tailored to the particular requirements of each application. The flexibility inherent in the printing process would facilitate the miniaturisation of delicate components, such as needles and microchannels, while concurrently enabling the fabrication of more robust reservoirs for field applications. Advancements of this nature have the potential to enhance the practical applicability of the device, thereby facilitating its transition from an experimental prototype to a scalable and efficient agricultural instrument.

How you can personalize for other needs

Although originally developed for agricultural applications, needles hold significant potential for adaptation across a range of sectors and can be customized to meet specific needs. In the domain of cosmetics, for instance, these devices have the potential to administer bioactive compounds, such as vitamins or antioxidants, directly into the deeper layers of the skin, thereby enhancing the efficacy of aesthetic treatments. In order to achieve this objective, customisations could involve the utilisation of biocompatible and even biodegradable materials that dissolve after application, thereby eliminating the need for removal.

In the environmental sector, needles have the potential to be utilized for the controlled release of beneficial microorganisms, nutrients, or bioremediation agents directly into soils or roots, thereby enhancing the processes of ecosystem restoration or soil enrichment in degraded areas. In this instance, the design of the needles would need to be adapted in terms of resistance and length to penetrate specific soil layers without compromising their structural integrity.

Another promising field is environmental and industrial diagnostics, where needles could be developed to collect fluid samples in diverse contexts, such as plant sap, biofilms in industrial systems, or even fluids in biotechnological materials. In order to achieve this objective, it is proposed that the surface of the needles be customized with specific coatings or integrated sensors, thereby enabling the immediate detection of compounds or organisms of interest.

Furthermore, in the domain of industrial biotechnology, needles have the potential to be utilised for the introduction of substrates or regulatory molecules into cell cultures or artificial tissues, thereby facilitating a more precise control over experimental conditions. Such adaptations may include geometric adjustments in design, variations in channel types, and the use of special materials such as metals or high-strength polymers to ensure reliable performance under different conditions.

In this way, needles stand out not only as an innovative tool in the fight against Greening disease but also as a versatile technological platform capable of addressing demands in diverse areas such as cosmetics, environmental applications, and biotechnology.

Steps on how to use the hardware

The injection hardware is designed to be simple and practical to use. First, the reservoir must be filled with the peptide solution to be applied to the plants, ensuring that it is properly sealed to prevent leaks. Next, the reservoir is attached to the body of the equipment.

With the reservoir attached, the user positions the applicator on the desired area of the plant. The device's design allows for stable and portable handling, facilitating transport between different areas and ensuring that the application is carried out without waste. After use, it is recommended to clean the reservoir and outlet points, ensuring the durability of the equipment and the effectiveness of future applications.