• machine learning
• opportunistic rainll estimation
• deep learning
• time series
• earth-satellite link

This thesis develops machine learning models to detect and estimate rainfall in real-time using raw Earth-to-satellite attenuation data. Accurate, real-time precipitation measurements are vital for areas like agriculture, flood control, and public safety. Classic techniques using rain gauges, weather radars, and satellite sensors aren't fulfilling the growing need for better accuracy, wider coverage and real-time estimation. A new method is to measure rain-caused attenuation in radio networks, especially Earth-to-satellite links in Ku and Ka bands. While terrestrial microwave links are well-studied, the Earth-to-satellite area is less known. This research addresses the shortcomings of classic methods in real-time by evaluating the capabilities of machine learning models. We tested models such as Gradient Boosted Machines and different deep learning setups, from neural networks to recurrent structures like GRU. We assessed model results using metrics like MAE, RMSE, and F1 Score, comparing them to traditional techniques. The goal is to offer a robust solution for estimating rainfall via Earth-to-satellite links. MBI s.r.l., a telecom company, supported this research, highlighting the practical applicability of the research.