AI in STEM: Inspiring Classroom Projects

The project integrates two complementary solutions for detecting and combating wildfires, which complement each other for an effective response. The solutions combine both hardware and software, merging physical sensors and artificial intelligence to optimize response time and accuracy.

1. The sensor-based system utilizes a flame sensor connected to an Arduino Uno board to detect infrared radiation emitted by flames. When fire is detected, the system immediately triggers a sound alarm and lights up an LED at a forest watchtower, alerting local response teams.
2. The AI-based smart solution employs AI to analyze images for fire patterns. Developed using the Pictoblox platform, the AI model is trained to recognize flames and distinguish between small and large-scale fires. This approach ensures precise detection and allows for real-time monitoring of fire progression. When the AI detects a fire, it sends a signal to the Arduino Uno, activating a relay that triggers a water pump to extinguish the fire. This automated response minimizes reaction time and enables immediate intervention without human assistance. Additionally, the system includes a remote monitoring mechanism via Adafruit IO and IFTTT. Upon fire detection, a wireless signal is transmitted to the IFTTT platform, which sends an automatic notification to the coordination team's smartphones, ensuring rapid communication.

Project phases

• In the initial phase, students conduct research on environmental sustainability, particularly on wildfires, their causes, and environmental impacts. They also explore key concepts related to Arduino, sensors, artificial intelligence, machine learning, and the Internet of Things (IoT), acquiring the necessary knowledge to develop the system.
• During the second phase, students create scenarios representing elements of the forest environment, such as trees and a watchtower. They design and build these elements using cardboard, integrating them into the test scenario. This stage helps them develop skills in digital design and manual prototyping.
• In the development phase, students assemble the electronic components (flame sensor, LED, relay, and mini water pump) connecting them to an Arduino Uno board. They programme the two systems - sensor-based and AI-based - using the Pictoblox programming environment. For the AI-based system, they utilise Pictoblox's Machine Learning tools to create and test models for fire detection in images. Additionally, students implement connectivity with IoT platforms (Adafruit IO and IFTTT), enabling remote system monitoring.
• During testing and optimisation, they fine-tune sensor sensitivity parameters and evaluate the AI’s accuracy in detecting fires. They also test the effectiveness of the automated fire suppression system, ensuring proper activation of the water pump in different scenarios.
• Finally, in the presentation and awareness phase, students document the process, create explanatory materials, and share the project with the school community and local entities.

The AI-driven component relies on a machine learning model trained to identify wildfires based on image analysis. The process consists of the following phases:

• Data collection: Representative images of various fire types are gathered, including small-scale flames and large-scale wildfires.
• Data classification: Images are categorized based on fire intensity and characteristics to enhance AI accuracy.
• Model training:The model is trained using labeled datasets to recognize different types of fires effectively.
• Model testing: The trained model undergoes validation to assess its precision in identifying and classifying wildfires.

Arduino programming and image recognition:
Following AI training, the Arduino Uno board is programmed using Pictoblox’s Machine Learning extensions. A webcam is integrated for real-time image recognition, classifying fires based on predefined categories.

Fire intensity-based response:
Depending on fire severity, the Arduino board executes a programmed action, activating a water pump to extinguish the flames. The water volume dispensed is determined by the intensity of the detected fire.

The Patella ferruginea limpet is one of the most endangered marine species in the Mediterranean, facing increasing threats due to the invasion of Rugulopteryx okamurae in the Strait of Gibraltar Natural Park. This invasive seaweed is altering marine ecosystems and endangering the limpet’s habitat. To address this issue, this project focuses on developing an autonomous monitoring system that leverages AI and IoT technologies to detect and assess the risk posed by the seaweed.

A Raspberry Pi, powered by a solar panel and battery, has been deployed in a coastal area where the limpet is found. Equipped with a camera, the system captures images periodically. These images are processed using a Machine Learning model trained in Teachable Machine, which identifies the presence of Patella Ferruginea and determines whether the invasive seaweed has encroached upon its habitat. The system operates in a 24-hour cycle to ensure continuous monitoring while optimizing energy consumption. At scheduled intervals, the Raspberry Pi and camera power on to capture an image of the habitat. The AI model analyzes the image to detect the presence of both the seaweed. If the system determines that the seaweed is invading the habitat, it sends an email alert with the captured image. If no threat is detected, the system shuts down to conserve energy and remains off for 24 hours before restarting the cycle.

Project phases

To develop a reliable AI model, a diverse dataset of images representing both healthy ("Limpet OK") and affected ("Limpet NOT OK") limpets was needed. As part of the project, students participated in a field trip to a coastal area, where they personally captured images of limpets in their natural environment. This hands-on experience allowed them to understand the importance of data collection in AI projects while also learning about marine ecosystems.

The collected images were then used to train the model in Teachable Machine, a user-friendly tool by Google. For training, the model was configured with 500 epochs, a batch size of 16, and a learning rate of 0.001 to ensure effective learning. The AI analyzed the images to recognize patterns distinguishing between healthy and affected limpets. After training, the model was tested with new images, successfully classifying limpets with high accuracy based on the characteristics it had learned. The results demonstrated a clear distinction between the two categories, confirming the reliability of the system in identifying threats to the species. The trained model was then exported in the TensorFlow Lite format for use on the Raspberry Pi.

Students also helped assemble and program the monitoring system. They tested the system in different conditions, refining the AI model based on new data. A custom enclosure to house all components was designed in Tinkercad and 3D-printed for assembly and outdoor protection.

The project was inspired by the daily challenges faced by people with memory disorders, such as elderly individuals with Alzheimer’s or other forms of dementia. The goal was to develop a technological solution that helps them recognize domestic environments and understand the function of each space through artificial intelligence. The class embarked on an interdisciplinary journey combining literature, neuroscience, biomedical engineering, and programming, allowing fifth-grade students to explore AI’s potential in a hands-on and meaningful way.
 

Project phases

• Memory, remembrance, and technology: The journey began with a reflection on memory, inspired by Marcel Proust’s famous Madeleine passage in À la recherche du temps perdu. Students discussed how memories are linked to sensory experiences and what happens when memory deteriorates. The class invited a social worker specializing in cognitive disorders, who explained how dementia patients perceive reality. Through simulations and real-life examples, students gained insight into the disorientation these individuals experience and how clear reference points can improve their quality of life.
• Artificial intelligence and decision trees: To understand how AI processes data, we collaborated with a biomedical engineer, who introduced the concept of decision trees. Students engaged in an unplugged activity called "Where are the keys?", simulating an object detection system. They worked in groups to create decision-making schemes that allowed an AI to find a missing object based on location, colour, shape, and proximity to other items. This exercise demonstrated the complexity of AI recognition and the need for proper training datasets.
• Unplugged activity on object detection, tracing, and retrieval: Before transitioning to programming, students participated in an unplugged activity to understand how AI can identify, track, and retrieve objects (see "Unplugged activity" for details)
• Creating the AI home assistant: After understanding machine learning mechanisms, students designed an AI assistant capable of recognizing domestic environments and helping individuals navigate their homes (see "Creating the AI assistant" for details) 

1. Object Detection (Recognizing objects): Students created a grid in the classroom and placed various everyday objects within it. Each student had a specific task: to find a given object without knowing its exact position, relying only on descriptions provided by others. This helped them understand how AI recognizes objects by analyzing patterns and visual characteristics.
2. Object Tracing (Tracking objects: After identifying an object, students had to track its movement through different environments. Some students were assigned to move or hide the object, while others had to record its movement, simulating how AI tracks moving objects.
3. Object Retrieval (Recovering objects): In the final phase, students were given a clue about where an object had been left. To locate it, they had to use reasoning based on the tracking data collected earlier.

This introduced the concept of pattern recognition and how AI retrieves information from incomplete datasets.

This activity clarified how object recognition systems work, helping students develop a tangible understanding of how AI processes images for detection and classification.

After understanding machine learning mechanisms, students designed an AI assistant capable of recognizing domestic environments and helping individuals navigate their homes.

1. Defining environments and objects: Students classified key home spaces.

• Bathroom → Objects: towels, toothbrush, soap
• Kitchen → Objects: plates, glasses, food
• Bedroom → Objects: pillows, blankets, lamp

They then mapped possible actions (e.g., "In the bathroom, you can wash your hands").

2. Training the AI model: Using Machine Learning for Kids, students uploaded images of rooms and objects. The model was trained to recognize spaces with high accuracy.

3. Building the interface in Scratch 3.0: Students programmed an interactive system in Scratch 3.0. The script

• captures an image from the user,
• uses the AI model for classification,
• provides a voice/text message explaining the recognised environment,
• suggests actions (e.g., "Here, you can wash your hands").

4. Testing and improvements: Students tested the system with real images and improved the dataset for better accuracy.

Visually impaired individuals face many challenges in areas ranging from shopping and walking on the street to checking money and selecting food. This project aims to enhance their independence and improve their quality of life by providing the essential information (product name and content, obstacle location, amount of money, fruit-vegetable freshness, etc.) in real time and through auditory feedback.

For this, students develop machine learning models in areas such as object recognition, text reading, and freshness analysis, and ensure that the system has a user-friendly interface.

Project components:

• Object recognition and product information reading: The USB camera connected to the Raspberry Pi 5 scans the product held by the user or the object in front of them. The image is processed and the object is classified with the help of OpenCV and similar libraries. The name of the detected object, and if it is a market product, its content or usage information, is conveyed to the user audibly.
• Money identification: The camera recognises the money held by the user. The machine learning model detects different banknotes and coins. The system audibly announces the identified amount.
• Obstacle detection and warning system: The camera continuously scans the environment. The position and distance of the obstacle are calculated. Warnings such as There is a wall in 3 steps or There is a staircase ahead are given to the user.
• Fruit/vegetable freshness analysis: Image processing techniques are used to analyse the colour and texture of fruits or vegetables. The user is informed whether they are fresh or not.
• Auditory Feedback (Text-to-Speech, TTS): All this information is converted into voice commands through TTS. It is delivered to the user via Bluetooth headphones or a speaker.