Land Surface Temperature Modelling over Geoclimatic Regions of Nigeria using Soft-computing Intelligence Techniques

This study used minimum (TN), maximum (TX), and mean (TM) land surface temperature data obtained from the Nigerian Meteorological Agency, Oshodi, Nigeria covering the period of thirty years (1984 - 2013) over twenty meteorological stations in Nigeria. These twenty stations were further divided into four climatic regions of Nigeria. The multilayer perceptron (MLP) and radial basis function (RBF) neural networks were employed to predict these three land surface temperatures (LST) series. The networks were created with twenty past years of the LST series as inputs to predict the next ten years (2003 - 2013) of the LST series such that 60% for training, 20% for validation, and 20% for testing the networks. The performance of outputs (the predicted ten years LST series) was evaluated using the standard statistics metrics. Performance analyses of the MLP and RBF models showed that their coefficient of efficiencies (NSE) have maximum values of 0.973 and 0.966 for prediction of TN in the Sahel region, 0.987 and 0.967 for TX in the Guinea Savannah region, 0.995 and 0.922 for TM in the Derived Savannah region, and 1.00 and 0.520 for TX in the Coastal region respectively. The error analysis using root mean square errors and coefficient of uncertainty also showed that the MLP network has lower values than those of the RBF network. Therefore, it can be concluded that soft-computing intelligence techniques especially multilayer perceptron networks are suitable for the prediction of land surface temperature series over Nigeria for practical purposes.