Previous work in the area of hand washing detection has hinted at the usefulness of atmospheric sensors for hand washing detection. Specifically, a humidity sensor can be used to measure nearby tap water flow using wearable devices. For this work, we expand on previous findings and a pre-existing dataset by recording 10 additional participants with a self-made open-source prototype recording device. We introduce an updated dataset with 20 participants instead of 10 participants, for which we make available IMU, humidity, temperature, and pressure measurements. The newly recorded participants conducted more complex background activities, which increased our dataset’s real-world relevance. Additionally, we show how to train an optimized deep-learning-based classifier on different parts of the combined dataset, improving on the previous study’s results, achieving significantly better F1 scores ( \(82\%\) instead of \(70\%\) ) on the pre-existing dataset. Furthermore, by leveraging a BIO-BANK semi-supervised pretrained model, we show that, unlike in previous work, the addition of humidity sensors to IMU data has a positive impact on the classification performance on the old and the new dataset, improving the F1-score on the combined dataset from \(60\%\) to \(68\%\) . All code and data are publicly available on GitHub.