In the week before the Sovereign 2 SilverStar Ski Marathon — two back-to-back 40 km days in interior British Columbia, classic on Saturday and skate on Sunday — I was analyzing training data from the build-up to the race. Each workout had a GPS track overlay on a topographic map plus line charts with heart rate, pace, and elevation. For interval workouts, I’d mentally retrace the session and try to visualize which sub-technique I was using on each segment, then compare that with the lines on the charts. The visualization looked something like this:

Four-minute intervals on a gradual incline, two-minute rest while returning to the start, repeat five times. This was on lower Amabilis Road, so I was doing V2, except for that thirty-second section around the sharp right turn plus another twenty seconds on the slight descent where I get in a few cycles of V2-alternate. Why is my heart rate spiking in V2-alternate?

The heart-rate strap I wear has an accelerometer that measures torso movement. I can use that data for running, where it gives me cadence, vertical oscillation, and ground contact time. When I’m skiing, it has no idea what’s going on. I could lie to the watch and tell it I was running, then try to interpret running-dynamics data for skiing, but it wouldn’t really tell me what I want to know: sub-technique.

Cross-country skiing is a fascinating and challenging sport with dynamic movement patterns. There are two main styles or techniques — classic and skating — and each has at least two main sub-techniques (or “gears”) that you use based on the terrain. With skating, for example, the slower and more powerful gear is the “V1” — Canadians and others call this “offset” — where you double-pole on alternating leg-skate pushes. This is primarily a climbing gear. “V2” or “one skate” is the gear where you push with your poles on every leg push, and you let the ski glide longer with each cycle. There are a handful of other sub-techniques: V2-alternate (two skate), free skate (no poling at all), double pole, tuck, and the dreaded herringbone or diagonal skate when you’re really struggling up a climb — or your body is shutting down after forty-five poorly paced kilometers, and you’re reduced to the winter equivalent of the zombie shuffle.

Becoming a skilled cross-country skier requires mastering each one of these sub-techniques — even the herringbone — and being able to switch between them dynamically based on the terrain. On typical rolling terrain you might enter a climb in V2; switch to V1 on the steepest section; back to V2 as you crest the hill; skip a pole plant with V2-alternate as the trail starts to descend; put in a few free skates before getting into a tuck for the main descent; emerge out of the tuck in a low free skate as the trail flattens; push with the poles every other leg push (V2-alternate) for most of the flat; increase the poling cycle to every leg push (V2) as friction starts to eat into the glide (you should have gone with the violet wax); then repeat all of this on the next climb.

What would be helpful is if your watch could track all of those transitions. They do this for run-tracking. The app uses the accelerometer in your watch, heart-rate strap, or other sensor to identify when you’re running versus walking — you only walked a bit, and only because you were choking on too big a bite of an energy bar (you were feeling good on pace, you promise). I bet they could even use the accelerometer to identify when you were reduced to a zombie shuffle at mile 153 of the Moab 240 on that boring straight desert road at 3 a.m., when the paint centerline started to wiggle — I wrote about that in The Hearth and the Col. COROS fortunately skipped that task in their app development backlog, so it just says I was “walking” (at 0.5 mph).

Cross-country skiing is a niche sport. There aren’t very many of us compared to running and other more common sports. It’s also a winter sport that requires expertly groomed snow — at least for skate skiing — using an expensive machine and wide trails that are cleared of downed trees and debris in the fall before the snow falls. (And climate change is really not helping.) The result is that cross-country skiers get no love from the wearables industry.

This project is my small attempt to change that.

Subscribe now

The pivot

I’ve actually been working on an app for cross-country skiing for a while. It’s called Glide, and the first version was more like Strava with skiing-specific features: a database of groomed trails, grooming and snow-condition updates, social features for sharing activities and comparing pace. Most skiers already do most of that on Strava itself, but Strava doesn’t help warn you that the corduroy is now velcro at 4 p.m. on a Thursday. Glide could fill that gap.

What Glide still couldn’t do was tell me anything about how I was skiing: the sub-techniques, the transitions. That gap stood out to me when I stumbled upon a paper showing it was possible to accurately detect sub-techniques using data from sensors similar to those in our sports watches and related accessories. Pousibet-Garrido et al. (2024),1 used inertial measurement units (IMUs) roughly comparable to the one inside my Garmin heart-rate strap, and classified three skating gears with about 90% accuracy across users and 98% for a single user with a neural network model. That was the moment I realized Glide didn’t need to be a better Strava for skiers. Glide needed to provide for skiers what the running-dynamics products provide runners. The initial plan for the app was about where and what to expect — corduroy, velcro, mashed potatoes, or ice. The pivot might still include that data, but the focus will be on how you skied in those conditions. How can you improve as a skier?

The research

There was more research on the question “can you classify XC ski sub-technique from inertial data?” than I anticipated. Stöggl et al. (2014)2 put a smartphone on a skier’s chest and showed that you could correctly classify skating gears 90% of the time with variable skiing (multiple gears) and 100% of the time when only a single gear was used. Seeberg et al. (2017)3 used seven IMUs plus a heart-rate sensor and GNSS, and correctly detected all three main classical sub-techniques — diagonal stride, double-poling with a kick, and double-poling — over 99% of the time. Pousibet-Garrido et al. (2024),4 mentioned above, mounted an IMU to each ski and correctly classified skating gears with a CNN-LSTM 98% of the time for an individual skier, and 90% of the time across skiers. Polo-Rodríguez et al. (2025)5 compared pressure-sensing insoles with inertial sensors for the same task and reached 94–99% accuracy with what they called a “minimal sensor setup.” For me, this was important provenance. I am interested in working with sensor data and building machine learning models that can run locally on mobile devices, but I am not in a position to pioneer the science of gear classification. However, that research is well established. There are still many open questions and opportunities, but the core question of whether it’s possible to detect ski sub-techniques using sensors similar to those available in the consumer market seems to be settled.

There’s another side to that question that’s important. Why would a skier want to track sub-techniques and transitions during their workouts? The answer is to turn the classifier into a coach. Can the data be used to assess how well the skier is performing? It turns out that researchers have been working on this as well. For example, Debertin et al. (2024)6 used principal component analysis on IMU traces of V2 skating and isolated the elemental components that an instructor or coach could use, including posture, lateral tilt, arm swing, ski set-down, leg push-off, and the gliding phase. This paper shows that data collected from these sensors can be used to identify both what you’re doing during the workout and how well you’re doing it. The output could be both “you skied V2 for 38% of that climb,” and “your ski set-down is late on your left, and your arm swing is shorter than it was last week.”

What’s missing is the consumer-grade product.

It’s all about the sensors

Learning more about the science was an important step forward for me with this project, but there’s a critical open question. Other than Stöggl et al. (2014)7 strapping a mobile phone to skiers, most of the research used custom, research-grade devices and software. Can this be done with consumer-grade hardware?

As an obsessive athlete (see The Hearth and the Col), I have a lot of gear, including three different sensor systems to track workouts and races. I use a mid-tier Garmin watch and HRM-Pro chest strap for shorter, intense workouts like intervals and strength training in the gym. I use a high-end COROS watch for long outdoor activities, especially for navigation, plus an arm-strap heart rate monitor and the POD sensor on my shoe that gives me running dynamics like cadence, stride length, and technique classification — running, walking, or the zombie shuffle (“walking”). I also have a small iPhone, which I take with me as a secondary navigation option and for all the other normal uses.

Each vendor makes a different decision about what an outside developer may do with the device. Those decisions are the difference between a buildable project and a non-starter.

Garmin’s Connect IQ SDK8 exposes accelerometer data through Toybox.Sensor. The catch is the sample rate: 25 Hz on my watch, in the documentation I found, with a potential 100 Hz buffer available for higher-end models (although it’s not clear if it can be utilized for what I need). Pousibet-Garrido et al. (2024),9 used a 112 Hz sampling rate for the IMUs on skis, and Polo-Rodríguez et al. (2025)10 used 100 Hz in their “minimal sensor setup.” It’s also not clear if I could tap into the stream coming from the HRM-Pro chest-strap, which Garmin uses for running dynamics. A typical pole-plant cycle is about a second long. At 25 Hz, that’s twenty-five samples per cycle. This is probably enough to count cycles (cadence), but almost certainly not enough to distinguish the waveform shape of V1 from V2. Both of these involve a pole plant followed by a leg push followed by glide, but with different timing and asymmetry. What this means is that Garmin is an option because the data is accessible, but it may not be the best one.

My COROS watch, in contrast, is top-of-the-line, and I have higher-accuracy heart rate and inertial sensors. Their API documentation11 shows that the API exposes activity-level data in the form of a fitUrl that points to a .fit file for each workout, or each leg of a multisport session. FIT is the industry-standard binary format for fitness data; what’s actually inside that file depends on what the device chooses to record. Based on COROS’s docs, it’s clear that the file has summary metrics, GPS, heart rate, and derived running metrics like cadence and vertical oscillation. However, it’s not clear if the raw accelerometer channels are available. This means I will have to inspect the file with a FIT SDK or a custom parser, and even then, I might find the data isn’t there. However, before I can access the file, the API is gated by an application process.12 Some developers in online forums complain of waiting for weeks or months, or not ever hearing back. I am not a company like Strava or TrainingPeaks, developing an interface for an established platform. I am just a solo developer developing an app for a niche user base. Ultimately, COROS is not going to be a viable platform for me to try to work with on this project.

The third option is the Apple Watch. I don’t have one, but I have other Apple products and am familiar with iOS development. The Apple Watch also gives a developer direct, documented programmatic access to the underlying inertial sensors. The Core Motion API hands you accelerometer, gyroscope, and magnetometer data through APIs without an application process. The gating happens at the App Store when the app ships to users. This means I can develop and test with off-the-shelf products and my existing Apple Developer account. Additionally, the Watch appears to have some of the best sensors. White et al. (2024)13 tested the Apple Watch accelerometer sampling at 100 Hz and found it to be as accurate as research-grade devices for the kinds of biomechanics work this project needs (more accurate than the other consumer devices they tested). While ~100 Hz is the rate used in the literature, the accelerometer in the watch was capable of up to 800 Hz with 2023 models.

Putting all of this together, two things are clear. This project is viable with consumer devices, and the Apple Watch is the best solution (at least for now). The price of admission is that I have to get another watch. New gear!

What’s next

It is summer, so it may seem like a bad time to develop and test a skiing application. However, dedicated cross-country skiers train on roller skis during the spring, summer, and fall. (Climate change is forcing this during some winters.) My focus during this period is on running, but roller skiing is a primary cross-training activity, and this gives me an excuse to hit the pavement more. Roller skiing is very similar to on-snow skiing, especially for skating, sharing the same sub-techniques (V1, V2, etc.), and it will be a good place to start collecting data and training models. The goal will be to have a version of Glide for the Apple Watch ready for skiers by winter after a full summer of training, data collection, machine learning, and iOS and watchOS development.

My plan for next Wednesday’s post is a deep dive into the research, including what IMU data looks like, how it can be used to detect skiing techniques, and why cross-user generalization is the hard problem.

I picked the Apple Watch for this project because it is the most viable option. However, I don’t actually have an Apple Watch yet. My data lives with COROS, Garmin, and TrainingPeaks. They have my heart rate, running form, weight history, training history, and sleep history. They even have a record of when I was “walking,” and the centerline started to wiggle.

All of that data went from sensor to Bluetooth, to watch, to phone, to the cloud. I am not an interesting target. No one really cares that I was actually doing the zombie shuffle (“walking”). But what if I were the CEO of an important company? Or a state representative? Or a famous athlete who is (not) hiding something? That’s a question I will return to in an upcoming Sunday essay and later in this Wednesday Wearables thread.

For now, see you next Wednesday.

Subscribe now