Implementation of Deep Neural Network Algorithm for Dual Polarization Radar Product Classification and Prediction

The evident success of the convolutional Neural Networks (CNN) in computer vision, speech recognition, image classification and prediction has been gaining immense and continuous relevance in medical imaging, communication systems, remote sensing, in military and defense, for prediction, estimation and security purposes. By harnessing the CNN model, a type of Deep Neural Network (DNN) algorithm, we can extensively utilize deep learning techniques to accurately recognize, predict and classify images from radar fields.In this work, the task is to design a heuristic CNN model for effective recognition, prediction and classification of storm observations from weather radar images. The feature extraction properties as well as the classification layers of this model would be trained, and validated with six different types of polarimetric radar field (observed storm).The data used was obtained from the ARMOR radar observed on the 11th of December 2021. I analyzed the model with a testing dataset from the six classes which were Reflectivity (Z), Differential Reflectivity (ZDR), De-aliased Doppler Velocity (DVEL), Doppler Velocity (VEL), Differential Phase (PHIDP), and Cross Correlation Ratio (RHOHV) when training was completed. This work was implemented with100 data samples for each of the radar fields of which 80 percent was to be used as training data set, while 20 percent of this dataset was generated for validation. In addition, 20 percent of the original data was labelled as the test datasets. Hence, atotal of 600 radar field images were used in this project.Upon dataset generation, the images were labeled and subdivided into training, test, and validation datasets. The task was completed by preprocessing images with MATLAB, where I resized, normalized and augmented the input image to fit the Model. The Google Collaboratory open source Integrated Development Environment (IDE) software was used for the execution of this work. Augmentation of both the train and validation dataset was carried out. I designed a CNN model with two convolutional layers which was alternated with a corresponding max-pool layer, a flatten layer, a drop out layer sandwiched by two dense layers.With a training accuracy of 98.00%, the test and validation accuracy were 97.5% and 98.9% respectively. Overall, the experimental results indicate that the algorithm could almost always classify, recognize and predict the different classes of radar storms observed as evident on the confusion matrix.