DIYALA JOURNAL OF ENGINEERING SCIENCES

Abstract

Remote sensing data is an important data source that can provide valuable information about urban expansion and urban land cover and land use changes at various scales. Due to the increasing spatiotemporal dimensions of the remote sensing data, traditional classification algorithms may not be able to classify such data. In this regards, two key issues should be taken into account: firstly, the challenges of the images fusion of the optical multi-source remote sensing data (MRSD) to seek the possibility of improvement in classification accuracy for urban change mapping. Secondly, monitoring and detecting the change and the interrelationship between land cover and land use within urban areas are spectrally and spatially complex.
Therefore, this paper aims to test and compare three classification algorithms (maximum likelihood (ML), decision trees (DT), and support vector machines (SVM)) for their ability to infer and extract urban land cover/land use across five different years using Landsat 5 TM, Landsat 7 ETM+ and ASTER images. Image pre-processing and post-processing were conducted on each scene along five different dates to obtain classification maps of Newcastle city, UK. Thereafter, the three aforementioned classifiers were used and applied on the combined data, which contained thirty-three bands in order to evaluate their effectiveness at separating urban land cover/land use types.
The classification approaches were implemented using ERDAS IMAGINE 2013 and coding by MATLAB. The results indicate that the overall accuracy of three classification maps using ML, DT, and SVM classifiers was 71.09%, 73.05% and 83.20%, respectively.
The classification maps present significant enhancement in the spectral and spatial resolution using optical MRSD compared to the one source of remote sensing data.