Comparative analysis of non-linear artificial neural networks and maximum likelihood algorithms in forest cover studies

Abstract

This study compares the level of uncertainty of a Back-Propagation Perceptron Network and the Maximum Likelihood classifiers (MLC) in the task of forest cover analysis. The input data comprises of bands 3, 4 and 5 of 2017 OLI Landsat image. Pixel grouping with these models was executed in Idrisi Selva using supervised technique. The degree of accuracy for each model was determined using 60 reference data. The results show that image classification with Non-linear Artificial Neural Networks algorithm (NANN), produce outputs with lower class weight RMSE of 0.02, and class weight RMSE of 0.14 was produced by Maximum Likelihood classifier. The overall accuracy of NANN (98.3%) is higher than that of MLC (80%). Standard errors at 85% confidence interval revealed NANN as a more effective statistical tool in separating forest from non-forest area. These indicate that misclassification of pixels occurred more with MLC than with NANN model. The comparison of RMSE values was possible because the same training data size, reference data and image were used for the different classifications.

Keywords: Back-Propagation, Multilayer Neural Networks, Maximum Likelihood classifier, Image classification and Misclassification