Statistical Approaches to Environmental and Atmospheric Influences on Fifth-generation Mobile Signal Propagation in Tropical Regions: A Critical Review

Abstract

Tropical and equatorial territories combine the most intense sustained rainfall on the planet with an accelerating rollout of fifth-generation mobile networks in the centimetre and millimetre wave bands. The statistical apparatus used to translate environmental and atmospheric measurements into propagation predictions, however, was assembled largely from temperate-latitude datasets and from fixed links considerably longer than the urban small cells that dominate contemporary deployment. This review critically evaluates how statistical methods have been applied to environmental and atmospheric parameters governing fifth-generation propagation in the tropics, and asks whether the resulting evidence supports the confidence routinely placed in it. Five method families are examined: the estimation of point rainfall rate distributions and the conversion of long integration-time records to the one-minute reference; the regression of specific attenuation on rainfall rate through power-law coefficients derived from drop size spectra; the spatial statistics embedded in path reduction factors, effective rainfall rate formulations and site diversity models; the correlational and regression treatment of water vapour, temperature, refractivity, wind and vegetation; and supervised learning applied to path loss and attenuation time series. Convergent evidence establishes that rainfall rates exceeded for one hundredth of one per cent of an average year in equatorial sites are roughly an order of magnitude above temperate values, and that standard international recommendations systematically misestimate attenuation on paths shorter than one kilometre. Beyond this, the evidence weakens sharply. Reported model rankings rest on single-year records from a small number of sites, are assessed with inconsistent error statistics and almost never carry uncertainty intervals, so apparent superiority is frequently indistinguishable from sampling variation. Studies of non-precipitation parameters report correlations that reverse sign between seasons and regression models that explain little of the observed variance, yet these are often interpreted causally. Machine learning results are constrained by small, spatially autocorrelated datasets, by evaluation protocols that permit optimistic bias, and in some cases by reliance on simulator output as ground truth. Progress requires coordinated multi-site campaigns, open data, formal uncertainty quantification and validation protocols that test transfer rather than fit.