Detection of weak snow layer on skis using radar and machine learning

Popular science presentation The Sknow BIA project included research on radar-antenna hardware for mountain and randonee skis, software development for radar signal analysis and interpretation, machine learning research that included deep dives in to the best fit, most recent developments in machine learning techniques of geo-data and complex natural behavior. Additionally, this project has performed great advances in interpreting and presenting advanced radar data to user benefit, and through user journey projects, user experience projects and testing of front-end solutions translate this to the commercial end-user. The project is un-concluded, as more time and resources is needed to conclude a commercial product. Hardware The hardware consists of a radar device mounted on a ski. This has gone through several iterations including: 1. Proof of concept: Radar development kit connected to a custom antenna embedded in a ski to prove the feasibility of the concept. 2. Prototype v1: Pre-production unit based on off the shelf radar chip. Center frequency of around 3GHz. Small and light enough to comfortably be able to mount on top of a ski. Custom antenna embedded in the ski. Abandoned due to concerns about future availability of chosen radar chip, prohibitive cost of said chip and issues with the embedded antenna. Åsnes provided input to ski behavior, ski mechanics and produced 30 ski prototypes with the antenna embedded. 3. Prototype v2: Pre-production unit based around a cheaper more future proof off the shelf radar chip. Center frequency of around 7GHz. SINTEF helped in verifying the viability of a redesigned antenna based on results from the former prototype. Antenna moved out of the ski and into the housing of the electronics. Allows for mounting of the device on off-the-shelf skis. This was very successful in dry snow and saw over 2m into the snow. 2m snow depth is sufficient since the weight of a skier will not induce an avalanche deeper than that. There was a problem with damp or wet snow where the signal was attenuated too much to penetrate through the zone of interest. 4. RR600: Custom radar with a substantially lower center frequency of around 600MHz to test that this lower frequency was sufficient to see through the avalanche zone of interest in wet snow. This has successfully shown that this frequency can see through wet snow for the whole avalanche zone of interest and still detect layers within the snowpack. Too large per today to mount on commercially available skis and needs further iterations. Future Hardware 5. Antenna redesign for use with RR600: Currently the antennas are Vivaldi which are not practical for mounting on skis. Our project partner, SINTEF, has designed new flat antennas suitable for mounting on a ski using the lower frequency. Antenna design is constrained by the width of a typical ski for use in the mountains. 6. Electronics Miniaturization: The current electronics will be miniaturized to fit on a ski and connected to the new flat antennas. Software The current software development includes: A. Think Outside Radar Processing (TORP): Conduct all the preliminary processing and signal attribute extraction. The preliminary processing and data conditioning is largely complete. Signal attribute extraction is continuing. B. Machine Learning Engine: Correlates the signal properties with the snow properties. Makes predictions of the snow layer properties. Training of the machine learning models is still under development. Investigations about the best machine learning model to use are currently undergoing. ML training by NGI based on a large database from Switzerland Snow and Avalanche institute (SLF) for patterns and trends in snow pack. Several synthetic snow pack models build, for ML project, to tie snow pack layers and radar traces. C. Internet-Phone Interface: Connects the machine learning engine to a user's mobile phone to send current data to the web, receive updates from other web data (weather, other ski data etc), inform the user of the current conditions and predicted avalanche risk. Web interface conceptual design is complete and wireframes and UI-Mockups have been established and tested with users. Coding implementation of the phone app has been delayed due to lack of financial means and time. Data Collection Radar data of snow has been collected with various prototypes and development kits throughout the length of the project. The focus was to collect a large and variable dataset of snow conditions, representing different snow depth, snow wetness/dryness, and different layer conditions. It has been collected primarily in Norway, but also includes data from various other geographical locations such as Chile, Italy, Austria, Switzerland and Finland. In fall of 2021 we deployed a fixed radar at Fonnbu research station in collaboration with NGI to measure snow and its properties throughout the winter season, including through the melting phase.

Within the frame of this BIA project a number of remarkable results have been achieved although the overall goal of a final product-stage version of an on-ski radar along with the according app has not yet been reached. We have come a long way towards it though with the design of an early version of the miniaturized ski sensor and additional, functional antenna and radar designs have been established and proven successful in practical field tests. A sled version of the mentioned developed radar sensors is operational and sold commercially. Moreover, over the course of the project a large amount of valuable snow data has been generated and combined with other large snow data bases for subsequent intensive analytics. This also allowed for the development of an operational software for processing and analysis of the snow data collected with the developed sensors. Additionally, the customer journey and web/app interface design has been completed and successfully tested with potential user groups.

Off-piste skiing have grown tremendously in popularity, and 1 in 3 skiers and snowboarders now venture away from prepared slopes. This increases exposure to steep and avalanche-prone terrain, so the number of skiers who have been harmed by avalanches is steadily growing. Current avalanche products are designed for rescue rather than prevention, so skiers must rely on their own assessments of the snow and terrain conditions to avoid risky areas. This method is often inadequate and consequently, 9 of 10 skier involved avalanches are triggered by the skiers themselves. Very often a persistent weak snow layer, unseen by the skiers, causes the avalanche and too often, loss of human life. Knowledge of avalanche danger is essential for safe skiing, especially in remote, high mountain areas. Regional forecasting, topographic maps and weather reports offer avalanche risk indicators, but these are based on extrapolations from few data points, and do not offer accurate location specific information. Our project is the engineering of a portable, light-weight, low-cost radar solution combined with other sensors: GPS, gyro, temperature and tilt, for skiers in the backcountry. It is the first mass-market device for real-time data acquisition and analysis of the snow conditions. It digitally measures a full snow profile including weak layers and snow depth, as well as slope gradient and applied forces. This technology will keep skiers constantly informed about the snow conditions under them, no matter their location. This data is not only useful to the skier, but to ski resorts, hydropower plants, militaries, insurance companies, climate researchers and others. Our 1st milestone will be the automation of weak snow layer detection using machine learning. Then we will look into integrating other sources of data (weather, gyro ,etc..) to provide localized avalanche risk information and look into communicating it in such a way that it does not encourage more risk-taking behavior.