Improving land use/land cover classification utilizing a hybrid method of decision trees and artificial neural networks

Monitoring the changes in land use/land cover (LU/LC) has become increasingly important for sustainable management of the natural resources to overcome the problems of haphazard development, deteriorating environmental quality, loss of fertile agricultural lands and destruction of wetlands. Comparable to the areas located at the northern part of the Nile Delta, EL-Manzala Lake and its surroundings are continuously subjected to numerous threats. These threats include converting parts of the lake into other land uses such as the fish ponds located to the south and south-east of the lake. The lake water is also subjected to pollution from the inflow of agriculture drainage water. Hence, the necessity arises to have a permanent monitoring strategy for this area to enable decision makers to develop plans to overcome these threats. The aim of this work is to develop a hybrid method that integrates decision trees (DT) and artificial neural networks (ANN) to increase the accuracy of classification in a selected area covering El-Manzala Lake and its surrounding. In this study, the DT was designed first in a manner that the classes can be separated and the DT results are then input into an ANN. Two satellites data were integrated as the input of the classification. The data included the reflectance of the Sentinel 2 satellite bands from the blue to the shortwave bands (discarding the red edge bands), and the emission of the thermal bands of the Landsat 8 satellite. The images of both satellites were acquired at the same date. Moreover, theses data were used in the format of band ratios and specific principal component analysis (PCAs). Utilizing this approach the overall accuracy of the classification could reach 98.6% while it could only reach 70.2 % when processing the data using the DT only and has been lower using ANN.