Rain activity detection using microwave link data with convolutional neural network

Repository of Masters thesis CNN implementation 2024/2025
author: Lukáš Kaleta

Included Jupyter notebooks showcase the data pre-processing and use of implemented CNN model

Data preprocessing

Done:

• fault values replaced with NaN
• long off times detected and removed from dataset
• large trsl steps removed or dealed with: (see cml 496)
• right approach for standardisation

01 training CNN on one CML

Goal:

• train existing CNN(1) on open sense cml and reference RADOLAN data from Germany(2).
• calculate WAA using pycomlink function from Schleiss 2013.

status:

• Prediction works with test loss around 0.2.
• Solved array size missmatch.
• choosing one cml, converting to torch Tensor.

Optimize learning:

• dropout rate: 0.4 is far to high, causes high learning curve ripple: set 0, later can be increased.
• learning rate lowered: 0.0001, learning is fast but convergs to high values.
• changed standardising: to min-max = 0-1, performance improved significantly!

02 training CNN on a dataset of 20 CMLs

status:

• Added more cmls to dataset. current: 20

TODO:

• sample shuffle: increases learning speed and precission.
• wet/dry 50/50 for faster learning and more accurate TPR/TNR results
• CNN threshold optimalization algorhythm, currently set to 0.5.
• improve the CNN architecture.

03 TODO after semestral thesis

• period of trsl == reference wet/dry. Meaning, for each trsl point there will be wet/dry flag predicted.
• Forward and backward memory implementation will be needed.
• This approach should bring better learning performance. For longer wet/dry periods there are ocasions, where the period is wet, but trsl shows rain pattern for only fraction of the period.

current CNN architecture (1):

Input (2 channels) → Convolution Block 1 → Convolution Block 2 → Convolution Block 3 → Convolution 5a → Convolution 5b → Flatten → Dense Layer 1 → Dropout 1 → Dense Layer 2 → Dropout 2 → Output Layer → Sigmoid Activation → Final Output (0 or 1).

Convolutional Part:

• 3 convolutional blocks with increasing filter sizes.
• Final 2 convolutional layers (conv5a and conv5b) to capture more complex patterns.
• ReLU activations after each convolution.
• no pooling implemented.

Fully Connected Part:

• Two fully connected layers with 64 neurons.
• Dropout applied after each fully connected layer to avoid overfitting.
• Final output produced through a single neuron with Sigmoid activation (ideal for binary classification).

sources:

(1) CML wet/dry using Pytorch: https://github.com/jpolz/cml_wd_pytorch/blob/main/wd_pytorch/train_cnn.ipynb
(2) Pycomlink: https://github.com/pycomlink/pycomlink
(3) JPolz: CML wet/dry using Tensorflow: https://github.com/jpolz/cnn_cml_wet-dry_example/blob/master/CNN_for_CML_example_nb.ipynb