Improving the accuracy of classification of multispectral Images using an anisotropic diffusion neural network algorithm (ADNNA) and machine learning SVM algorithm

Abstract

To improve image classification accuracy, one common practice is to add specific information
like image texture, DEM and different indices to satellite image datasets. Here, an attempt has
been made to include outputs of Anisotropic Diffusion Neural Network algorithm (ADNNA) to
satellite image datasets during classification stages. The unsupervised multi-level neural
network (or anisotropic diffusion neural network) modifies the pixel values of the input image
by consecutive weighted averaging with neighboring pixels. It performs simultaneous
modification of all input multi-spectral image at five level of scale/resolution. The algorithm
can process both spectral information of the image and textural details resulting from wavelet
transformation simultaneously in a multi-scale representation. To perform this task Landsat
8 Surface reflectance images pertaining to Miandoab region have classified with and without
adding outputs of ADNNA (five levels) by support vector machine (SVM) algorithm. Different
dataset was created. First dataset was a composite of bands of original images only, second
dataset was a composite of different outputs of anisotropic diffusion neural network algorithm
only and third dataset contains band of original image and each output of anisotropic diffusion
neural network algorithm and the last had bands of original image together with all outputs
of anisotropic diffusion neural network algorithm. SVM classification algorithm was used to
classify all datasets separately with the same training site input. Classification accuracy was
performing through Kappa coefficient of agreement. Results show that the highest kappa
coefficient of agreement in classifying the Landsat 8 image is with level 2 of the ADNNA pluse
original image bands with approximately 86% compare to other datasets (original images,
all 5 level ADNNA outputs, other levels pluse original images). It is concluded that ADNNA
outputs can improve classification results effectively.