Using 1D convolutional neural networks (CNNs) for rain event detection in commercial microwave link (CML) data is a concept described in our paper
'Rain event detection in commercial microwave link attenuation data using convolutional neural networks',
which is currently in discussion. The purpose of this repository is to give a full description of the model setup using the Keras API with a Tensorflow back-end and a small example data set to illustrate our results. CMLs can be used to derive rainfall information by exploiting the close to linear relationship between the attenuation caused by rainfall and the path averaged rain rate along the link path. Rain event detection in the CML attenuation time-series is a binary classification task. Therefore we use the CNN as a binary classifier. A sample of time-series data consists of 180 minutes of instantaneously measured signal levels from one CML. The first 120 minutes are used as a reference to previous behavior and the last 60 minutes are the period where attenuation by rainfall has to be detected. To label the samples we use the gauge adjusted radar based rainfall product RADOLAN-RW provided by DWD.
The python implementation of the model is shown in this notebook. More details soon...
For the newest model version use
We train our model using 4 months of data from 400 randomly chosen CMLs distributed over entire Germany. Afterwards the model is validated using different months of data from all 3904 available CMLs. The number of samples is 70,000 for training and more than $ 4x10^{6} $ for validation.
Since the full data set, that was used in (link to publication-coming soon) is not publicly available in its full extent, we provide a small example data set. It contains raw data from 40 CMLs with modified CML locations. The time series contains two weeks in September 2018. The pre-processing of the raw data, the classification through the CNN and a comparison to a reference method is contained in this notebook.
The full python programming environment needed to set up the analysis can be installed from the environment.yml file. The most important toolboxes we used are:
-
Hydrological application of CML derived rainfall maps in alpine regions
-
Rain event detection methods