Machine Learning Techniques for Evaluating Disdrometer-Derived Raindrop Measurements Over Radio Links

• Main Authors: Ramatladi, Tsietsi, Alonge, Akintunde
• Format: Conference Proceeding
• Language: English
• Subjects: decision tree regression (DTR); Forestry; k-Nearest Neighbors regression (KNNR); Machine learning; Predictive models; Radio links; Rain; raindrop size distribution (DSD); random forest regression (RFR); Satellite broadcasting; Training; Wireless communication
• ISSN: 2153-0033
• DOI: 10.1109/AFRICON55910.2023.10293561

Rainfall drop size distribution (DSD) is an essential parameter for designing and operating radio links as found in numerous terrestrial and satellite systems. Utilizing machine learning techniques to analyze the number of raindrops within specific diameter bins can enhance DSD computations and provide valuable insights into other rainfall characteristics. This study investigates three regression models - decision tree regression (DTR), random forest regression (RFR), and k-nearest neighbours regression (KNNR) - as tools for evaluating raindrop measurements from the Joss-Waldvogel disdrometer (JWD). High-resolution rainfall measurements obtained from Durban, South Africa, are used to train and test regression models. The three proposed techniques accurately predicted raindrop presence across designated channels. However, the KNNR model performs better compared to other models, for all categories of investigated rainfall regimes. This research demonstrates the effectiveness of machine learning (ML) models in predicting and replicating disdrometer measurements over radio links. The findings in this study can serve as valuable input for radio link design over sub-tropical areas like Durban, where rainfall is highly variable.