A Machine Learning Approach for the Classification of Wet and Dry Periods Using Commercial Microwave Link Data

Rainfall estimation is essential for many reasons. The lack of rainfall monitoring and estimation infrastructure in some parts of the world, including sub-Saharan Africa, is a serious problem with severe consequences. For this reason, the possibility to use mobile network infrastructure, which are in continuous expansion all over the world, to monitor and estimate rainfall has been widely investigated during the last 2 decades. One of the critical tasks in the process of rainfall estimation using microwave links is the detection of rainy periods. Many methods have been proposed to realize this task in the literature, with satisfaction and underlying uncertainties resulting from mobile networks’ infrastructures and modeling assumptions. Machine learning has been proven to be very powerful for classification tasks and patterns recognition in high-dimensional data, as evidenced by many applications in various fields of science and engineering. Deep learning in particular has been recently used to improve the wet–dry periods classification task in rainfall estimation using commercial microwave links. In this paper, we propose to use other machine learning techniques for the classification of dry and wet periods using commercial microwave link data. We evaluate and compare the ability of single machine learning classifiers, ensemble machine learning classifiers, and clustering algorithms to successfully realize this task. We consider a sample of four microwave links located in the Netherlands (Western Europe), for which minimum and maximum received signal levels are recorded for each 15 min period from 30/05/2012 to 01/09/2012. In general, the considered algorithms are performant for this classification task, particularly for ensemble machine learning classifiers, among which an ensemble voting classifier emerges as a good first choice with an overall accuracy, and precision of wet periods detection very close to 100%. Its precision of dry periods’ detection is also close to 100% for the majority of the considered microwave links. These results show that machine learning classifiers are very effective for the classification of wet and dry periods for rainfall measurement using microwave links, and outperform some empirical and classical model-based methods available in the literature.