Home

The edition 2024 of the Smart-weather-station challenge has come to an end. Stay tuned if you want to get the opportunity to participate to a next edition.

Welcome to the Next-Gen tinyML Smart Weather Station Challenge! On this website you will find general information and description about the challenge.

Introduction

The "Next-Gen tinyML Smart Weather Station" competition is a challenge aimed at inspiring and promoting the development of innovative, energy-efficient, and cost-effective smart weather stations using Tiny Machine Learning (tinyML) technology. Participants from diverse backgrounds are invited to design, build, and deploy weather stations that can accurately measure and report real-time environmental data (rain and wind), as well as temperature, humidity, pressure, and/or air quality. The competition encourages the use of low-power, low-cost hardware, and software solutions that are capable of processing data locally and can operate autonomously for extended periods. The competition is expected to foster collaboration, creativity, and practical solutions that can contribute to improving weather monitoring thereby enhancing community resilience and adaptation to climate change.

Problem statement

The goal of this challenge is to create a low-cost, low-power, reliable, accurate, easy to install and maintain weather station, with no mechanical moving parts for measuring all weather conditions with a focus on rain and wind, based on ultra-low power machine learning at the edge, that can be deployed locally. This weather station could be deployed in a farm, for example, to provide local conditions and assist farmers in deciding when to plant crops.

To get you off with a flying start, the Swiss Technology Innovation Center (CSEM) will provide a dataset acquired by their winning Weather Station Aurora from last year’s challenge. This dataset contains microphone recordings and environmental sensor data (temperature, humidity and pressure). In addition, it will come with ground truth information from a mechanical weather station, which gives information about the wind and rain intensity.

Suggested process

1. Start with the dataset provided by the tech partner CSEM, to collect local rain and wind audio and other types of sensor measurements with an embedded device without any moving parts. In addition, you are also invited to come up with your own local dataset.
2. Develop a tinyML model to derive rain and wind intensity from sound measurements of rain and wind.
3. Optimize the memory footprint and the power consumption of the tinyML model.
4. Deploy the tinyML model on an embedded device in the field and measure how well the model performs in real-life.
5. Document the process, write a report and publish a video of the prototype.

Final goal

Develop a fully functioning rugged and cost-effective weather station based on a single device able to detect all weather parameters including TPH (temperature, pressure and humidity), wind and rain with no moving parts. Participants can get off with a flying start, using the dataset from the Smart Weather Station Aurora which will be made available by CSEM. The architecture of Aurora is depicted in the figure below. Exploiting this dataset, participants are challenged to go further, and to propose their own robust and unique embedded solution for a Smart Weather Station that can be deployed in the field.

About CSEM

Founded in 1984 and headquartered in Neuchâtel, CSEM is an internationally recognized innovation specialist with over 550 employees across six locations in Switzerland and more than 200 registered patents. We develop disruptive technologies with a high societal impact in the fields of precision manufacturing, digitalization, ultra-low-power electronics, optical elements, AI, and sustainable energy. We then transfer these innovations to industry partners in a variety of sectors, including renewable energy, healthcare, watchmaking, and aerospace, or encourage start-up creations. As a public-private, non-profit organization, our mission is to support the innovation of Swiss companies and strengthen the economy through ongoing collaboration with leading universities, research institutes, and industrial partners. Check out our website: here.

The evaluation will be performed by benchmarking your designed tinyML model against the models from the other participants, on the test set of Aurora, which will be released at the end of the challenge. Participants should provide the proposed tinyML model, as well as the Python script that reads in the test dataset, feeds it to the tinyML model, and outputs the evaluation metrics.

The ranking will be done based on the combination of the evaluation metrics listed below:

• Accuracy: the proposed tinyML model should estimate the intensity of wind and rain with a high accuracy (analyzed through precision and recall metrics). In addition, the output should be as fine-grained as possible, to closely match the intensity from the groundtruth weather station.
• Memory footprint: the proposed tinyML model should be optimized for both FLASH and RAM usage.
• Energy consumption: the proposed tinyML model should be optimized for energy consumption, to enable a long battery lifetime of the weather station.
• Latency: the proposed tinyML model should be optimized for energy consumption, to enable a long battery lifetime of the weather station.
• Prototype: Participants should build a prototype that can execute local inference of the proposed tinyML model on a resource-constrained device of choice.
• Documentation and Code Quality: Participants should provide clear documentation detailing their approach, tinyML model architecture, optimization techniques, and scalability considerations. The submitted code should be well-organized, modular, and easy to understand, with appropriate comments and explanations.

A detailed overview of the quantitative metrics that will be used for the evaluation will be shared with the participants throughout the challenge.

If you are interrested to participate, please register via the following link. Registrations are open to everyone from January 24, 2024 until May 10, 2024.

Proposal checkpoint

Your team is tasked with submitting a comprehensive proposal paper that outlines the functioning and structure of your model. The proposal should include detailed explanations of how your model operates and its underlying architecture. Additionally, it should address the expected performance metrics of the model and identify potential challenges that might be encountered during the development process.
Deadline: Monday February 26, 2024

Q&A session

We will organize a Q&A session with the organizers of the challenge. In this session, you can ask all your technical questions related to the challenge. We expect that at least one member per team attends the Q&A session.
Will be organized in the Week of March 25, 2024

Model checkpoint

For this milestone, your team will be required to present a progress report. This report should provide an overview of the current stage of development, including the results obtained so far. Explain the findings and outcomes from the ongoing work and detail any advancements made in the model's performance. Furthermore, outline the current challenges your team is facing during this stage of development.
Deadline: Monday April 15, 2024

Final submission

In this final milestone, your team must submit a comprehensive paper summarizing the architecture and functioning of the fully developed model. The paper should provide a complete description of how the model operates and its design details. Additionally, it should offer insights into the training and validation procedures employed during the model's development.
Deadline: Monday May 20, 2024

We are offering an opportunity to gather in a collaborative manner interesting insights on the promising field of tinyML to build a low-cost and robust Smart Weather Station. At the end of the challenge, we might reach out the teams that particularly stand out to initiate collaboration for further development of the project.