FEED-FORWARD AND CASCADE BACK PROPAGATION ARTIFICIAL NEURAL NETWORK MODELS FOR PREDICTING AEROSOL OPTICAL DEPTH IN ILORIN-NIGERIA

Abstract

This study presents feed forward and cascade back propagation Artificial Neural Networks (ANNs) approach to predict daily Aerosol Optical Depth (AOD) in Ilorin-Nigeria  using Bayesian regularization (trainbr) and Levenberg– Marquardt (trainlm)) algorithms. The daily AOD data for ten years (2007-2017) was collected from Aerosol Robotic Network (AERONET) and was modelled as a function of meteorological data (visibility and relative humidity) downloaded from NOAA-NCDC (National Oceanic Atmospheric Administration-National Climate Data Centre). The data linearity was checked, and the network was trained and tested using 70% and 30%. The numbers of neurons in the hidden layer were varied 3-3, 6-6, 7-7, 15-15, and 20-20 using tansigmoidal transfer function. In each adjustment, the networks were trained 15 times and the results that represent the best output were recorded. The combination which has the highest coefficient of determination R² and the lowest errors were finally chosen. Results indicate that visibility and relative humidity in all the two networks and algorithms using both the overall and seasonal data can be used to explain at least 71% of the variation in AOD. The ANN has shown high accuracy on the overall (January-December) compared to the Harmattan and summer seasons data. The result also showed that, the feed forward network performed better than the cascade network. It further revealed that both feed forward and cascade provide lowest error (lower RMSE) and faster convergent speed (lower epoch) when train with Levenberg–Marquardt algorithm (trainlm) than the Bayesian regularizations algorithm