4D-GWR: geographically, altitudinal, and temporally weighted regression

Taşyürek M., Çelik M.

NEURAL COMPUTING AND APPLICATIONS, vol.34, pp.14777-14791, 2022 (SCI-Expanded)


Geographically weighted regression (GWR) models varying relationships and is a local spatial regression approach. It has been used in several application domains, such as meteorology, environmental management, ecology, etc. The datasets collected in these applications include spatial (latitude, longitude), altitudinal, and temporal nonstationarities and so the values of parameters that are found in such datasets change over space and time. In the literature, several GWR models have been proposed to handle such datasets. However, these methods do not consider spatial, altitudinal, and temporal nonstationarity in the datasets simultaneously. This study deals with developing a GWR technique, 4D-GWR, to capture spatial, altitudinal, and temporal nonstationarities to improve the prediction accuracy. In addition, a new parameter estimation strategy, which utilizes n-dimension golden section search algorithm, is proposed to estimate parameters of 4D-GWR approach. Experimental evaluations were conducted to compare the proposed 4D-GWR approach with the classical approaches of GWR (2D-GWR), GAWR, GTWR, and GWANN using real-time and synthetic meteorological datasets. The results showed that 4D-GWR approach outperforms other approaches in terms of RMSE between the predicted and actual air temperatures, runtime, and dataset size. Experiments showed that GWANN could not handle more than 15,000 observation points, 2D-GWR, GAWR, and GTWR could not handle more than 40,000 observation points and, in contrast, the 4D-GWR could handle all dataset sizes, successfully. When the actual and predicted air temperature values are compared, the highest correlation of 0.95 was obtained by 4D-GWR and the lowest correlation of 0.77 was obtained by GWANN.