Water usage in residential buildings highly impacts the overall urban demand. Following the release of the first public high-resolution smart meter data at the fixture level, we envision that extending non-intrusive load monitoring to water disaggregation will avoid installing one smart meter per appliance while still providing users with insights into their consumption habits, which eventually promote water savings. The interest in deep learning in the energy sector has increased over the years, motivated by superior accuracy in real-world settings with many appliances. In light of this, the work aims to explore the effectiveness of deep neural networks in disaggregating water usage data by proposing a UNet architecture for near real-time multi-appliance water disaggregation. Experiments on various time resolutions show interesting results. Further qualitative analysis highlights the challenge posed by data sparsity in water end-use datasets and suggests possible research directions.