In this article we present an alternative method for extrapolation of Land Surface Temperature (LST) by means of Artificial Neural Networks (ANNs) based on positional variables (UTM coordinates and altitude), temperature and average air relative humidity. The study region was the Rio dos Sinos Hydrographic Basin (RSHB), Rio Grande do Sul, Brazil. For ANN training we used an NOAA-14/AVHRR satellite thermal image, with pixels size 1 x 1 km, with known information of LST on January 29, 2003. Various settings were tested in ANN training step, the one that presented the best performance was composed of only one intermediate layer (with 4 neurons and logistic sigmoid activation function). The trained network was validated with 2 simulations: in the first simulation we extrapolate the LST values of April 11, 2003 and in the second simulation we extrapolate LST values of October 15, 2003. The results of the simulations were compared with Split Window (SW) algorithm and the average discrepancies found between both models were of -0.30° C and 0.26° C, respectively, of April 11, 2003 and October 15, 2003. A strong correlation was found between both models with R 2 values exceeding 0.93 and statistically we checked that there was no difference between the LST averages values obtained by ANN and SW for 5% significance level.

International Journal of Sciences