Artificial Neural Network Model for Prediction of Land Surface Temperature from Emissivity & NDVI Images |
Paper ID : 1103-SMPR-FULL |
Authors: |
Farinaz Farhanj * gisha Ave, 15th street, #14 |
Abstract: |
The urban air temperature is gradually rising in all cities in the world. One of the possible causes is the reduction in the greenery area in cities. Hence LST is primarily depends on the amount and variable of vegetation at the surface and different types of land cover. An artificial neural network model has been developed for the estimation of land surface temperature (LST) using NDVI and emissivity parameters. Bands 3 (RED) and 4 (NIR) were used to calculate the Normalized Differenced Vegetation Index (NDVI). Different types of land cover have different values of emissivity. An Emissivity image is developed using the classified image and the NDVI image. Emissivity values are given as 0.96 for bare soil, 0.98 for vegetation, 0.99 for thick vegetation and 1 for water. To compose the ANN training structure we chose randomly pixels from the Landsat satellite image. An artificial neural network model has been developed in MATLAB with NDVI and emissivity of these pixels as input parameters and LST value of them as output parameter (for the ANN supervised training we used LST information coming from of Landsat thermal band). The ANN architecture consisted of three layers: the input layer, hidden layer (with 20 neurons) and output layer (with 1 neuron). Transfer function of the output layer and hidden layer that we used was the linear and tangent sigmoid respectively. The ANNs were trained with the Levenberg-Marquardt algorithm (LM). We trained the ANN in inputs and outputs pairs, for each input supplied to the network there is an expected output which is also provided for the training. The network produces an output answer that is compared with the desired output. The mean squared error (MSE) is calculated between expected and actual outputs. The weights and biases were optimized by using a defined genetic algorithm with the aim of error minimization. Population size, number of generations, crossover rate, mutation rate, mutation function and recombination function were chosen 50, 1000, 0.7, 0.0175, Uniform and Heuristic respectively. The goodness of fit statistics shows that R2 value and MSE are 0.6352, and 10.7638 respectively. Hence by using the artificial neural network model the LST can be predicted easily from NDVI and emissivity images. |
Keywords: |
land surface temperature, NDVI, Emissivity, Artificial neural network |
Status : Paper Accepted (Poster Presentation) |