Survey of different data-intelligent modeling strategies for forecasting air temperature using geographic information as model predictors

Air temperature modelling is a paramount task for practical applications such as agricultural production, designing energy-efficient buildings, harnessing of solar energy, health-risk assessments, and weather prediction. This paper entails the design and application of data-intelligent models for air temperature estimation without climate-based inputs, where only the geographic factors (i.e., latitude, longitude, altitude, & periodicity or the monthly cycle) are used in the model design procedure performed for a large spatial study region of Madhya Pradesh, central India. The evaluated data-intelligent models considered are: generalized regression neural network (GRNN), multivariate adaptive regression splines (MARS), random forest (RF), and extreme learning machines (ELM), where the forecasted results are cross-validated independently at 11 sparsely distributed sites. Observed and forecasted temperature is benchmarked with the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe's coefficient (E), Legates & McCabe's Index (LMI), and the spatially-represented temperature maps. In accordance with statistical metrics, the temperature forecasting accuracy of the GRNN model exceeds that of the MARS, RF and ELM models, as did the overall areal-averaged results for all tested sites. In terms of the global performance indicator (GPI; as a universal metric combining the expanded uncertainty, U₉₅ and t-statistic at 95% confidence interval with conventional metrics, bias error, R², RMSE) providing a complete assessment of the site-averaged results, the GRNN model yielded a GPI = 0.0181 vs. 0.0451, 0.1461 and 0.6736 for the MARS, RF and ELM models, respectively, which concurred with deductions made using U₉₅ and t-statistic. Spatial maps for the cool winter, hot summer and monsoon seasons also confirmed the preciseness of the GRNN model, as did the 12-monthly average annual maps, and the inter-model evaluation of the most accurate and the least accurate sites using Taylor diagrams comparing the RMSE-centered difference and the correlations with observed data. In accordance with the results, the study ascertains that the GRNN model was a qualified data-intelligent tool for temperature estimation without a need for climate-based inputs, at least in the present investigation, and this model can be explored for its utility in energy management, building and construction, agriculture, heatwave studies, health and other socio-economic areas, particularly in data-sparse regions where only geographic and topographic factors are utilized for temperature forecasting.