A Stacked Ensemble Machine Learning Model for Predicting Ionospheric Total Electron Content Over Uganda
| dc.contributor.author | Adriko, Norbert Bakole | |
| dc.date.accessioned | 2026-09-15T02:03:18Z | |
| dc.date.available | 2026-09-15T02:03:18Z | |
| dc.date.issued | 2026-09-15 | |
| dc.description | A research and capacity-building project led by Busitema University's Department of Physics (Faculty of Science and Education), Uganda, aimed at developing an operational space weather based early warning system for eastern uganda. The project also aims to develop human capital by training masters and phd students in the field. its sponsored by Uganda's Science, Technology & Innovation Secretariat (Office of the President). | |
| dc.description.abstract | Ionospheric delay, which is directly determined from Total Electron Content (TEC) is the largest source of error in GPS positioning. Because many GPS receivers widely used in Uganda lack the capacity to measure this delay directly, they must rely on models to correct for it. The models currently available are global and therefore capture neither the steep latitudinal gradient of the equatorial ionization anomaly or the day-to-day variability typical of low-latitude regions such as Uganda, which creates a need for a local model. This study first characterised the behaviour of TEC over Uganda and then developed a regional TEC model using a stacked machine learning ensemble trained on 26 years of Global Ionosphere Map data and six months of TEC observations from the Continuously Operating Reference Station (CORS) network in Uganda. The local correction stage relies on observation-derived residual lags, which bounds its evaluation to the station density and time span sampled here. The characterisation revealed a post-noon maximum between 13:00 and 17:00 local time, with night-time values remaining above 5 TECU. A semi-annual anomaly was also evident with the March monthly mean exceeding the June–July mean by 66%. The storm-time response proved to be a latitudinal redistribution rather than a network-wide change during the November 2025 storm: TEC at the northernmost station (ABON) rose markedly above its quiet-day baseline, while the southernmost station (KABA) tracked its own baseline to within 1 TECU. Global Ionospheric Map observations and IRI-2020 carried a mean bias of +18.0 TECU and +27.9 TECU respectively when compared against the ground truth CORS network observations. On the test period, the full ensemble achieved a RMSE = 3.970 TECU, R2 = 0.958 and a Skill Score of +0.724 against the climatological mean. This correlates with an 80% reduction in RMSE relative to the background field at the same locations thereby reducing ranging error of L1 from 3.26 m (background error) to 0.64 m (ensemble prediction). Spatial performance showed a strong trough-to-crest TEC latitudinal transition; leave-one-station-out cross-validation confirmed the model generalises to unseen locations, though with degraded accuracy (RMSE rising from 3.97 to 6.41 TECU) relative to stations seen during training. These observations suggest that the ensemble is a useful tool for ionospheric delay correction over Uganda. | |
| dc.description.sponsorship | Space Weather Based Early Warning Tool Project | |
| dc.identifier.citation | Adriko, N. B. (2026). A stacked ensemble machine learning model for predicting ionospheric total electron content over Uganda [Masters Thesis]. Busitema University. | |
| dc.identifier.uri | https://bdears.busitema.ac.ug/handle/123456789/9516 | |
| dc.language.iso | en | |
| dc.publisher | Busitema University | |
| dc.title | A Stacked Ensemble Machine Learning Model for Predicting Ionospheric Total Electron Content Over Uganda | |
| dc.type | Thesis |