SCORE project partners from MBI and UNIPI have presented a paper at the GLOBECOM 2024. It is available in open access.
Abstract
Rainfall estimation, a key task for enabling sustainable and intelligent living, has traditionally been carried out using dedicated devices. Recently, machine learning (ML) applications exploiting opportunistic satellite-to-earth microwave links (SMLs) have gained traction. However, significant deep learning implementations to perform end-to-end regression from raw signal-to-noise ratio (SNR) to rainfall estimation remain underexplored. Leveraging a proprietary satellite receiver network, we address this shortcoming by building a conditional ensemble neural network (a deep learning classifier-regressor chain) and comparing its performance, both in terms of detection and regression, with a state-of-the-art power-law (PL)-based algorithm. Our neural network shows significant gains in both tasks. Regarding the classification task, the gains are marginal: approximately a 3.5%-increase in both accuracy and F1-score. On the other hand, for the regression task, the performance gains are more substantial: a 34%-decrease in normalized mean absolute error (NMAE) and a 70%-decrease in mean absolute percentage error (MAPE) on event-based cumulative precipitation predictions.
References
Title: Opportunistic sensing with satellite communications using a conditional ensemble neural network
Authors: Giovanni Scognamiglio, Giacomo Bacci, Attillio Vaccaro, Andrea Rucci, Fabiola Sapienza, and Filippo Giannetti
Cite as: Scognamiglio, G., Rucci, A., Vaccaro, A., Bacci, G., Sapienza, F., & Giannetti, F. (2024, décembre 31). Opportunistic sensing with satellite communications using a conditional ensemble neural network. https://doi.org/10.5281/zenodo.15967929
Find all SCORE’s publications on this page.
