Groundwater storages are usually inaccessible and therefore their surveillance can become challenging. As a complement to traditional methods, seismic noise analysis was suggested to monitor ground water storage (Lecocq et al. 2017). Our site is the Fourbanne karstic aquifer, monitored since 2014. The underground conduit is accessible through a drilled well and is instrumented by two 3-components seismometers and a hydrological probe. We present a new approach, based on the machine learning random forest (RF) algorithm and continuous seismic records, to find signals corresponding to flooding and predict the underground river water level. The method is based on the computation on a sliding window of seismic signal features (waveform, spectral and spectrogram features). The first results indicate that the RF algorithm is capable of accurately predicting the water level in the conduit, with a mean absolute error not exceeding 5%. This a first promising outcomes for the remote study of water circulation using seismic waves.