Vahid Sadeghi; Hamid Enayati; Hamid Ebadi
Abstract
Analyzing multi-temporal remotelysensed images is an effective technique for detecting land useand land cover changes in urban areas. Apart from thetechnique used to detect the changes, the features space has an enormous impact on the accuracy of the results. Achieving satisfactory results in detecting ...
Read More
Analyzing multi-temporal remotelysensed images is an effective technique for detecting land useand land cover changes in urban areas. Apart from thetechnique used to detect the changes, the features space has an enormous impact on the accuracy of the results. Achieving satisfactory results in detecting changes inurban areasrequires the use of optimal spectral and spatial features (texture). Although global search is the only guarantees of achieving the optimal set of features, but it is a very timely and impractical process in practice. Data reduction techniquessuch as PCA considers the independence of the data tofind a smaller set of variables with less redundancy withoutintending to improve the CD accuracy. Difficulty in setting thebest threshold for JM distance in Separability Analysis Algorithm (SAA)reduces its efficiency. The main purpose of this paper is to select the optimaltextural and spectral features to enhance the CD accuracy usinggenetic algorithms (GA) and Bayesian classifier. To investigate the effectivenessof the proposed tecknique, a case study using IRS-P6and GeoEye1 satellite imagery taken from Sahand New Town (Northwest ofIran on July 15, 2006, andSeptember 1, 2013) was performed. All of the aforementioned methods of feature selection (PCA, SAA and proposed GA-based method) were implemented in MATLABR2013a. The results show that, textural features provides a complementary sourceof data for CD in urban areas. The results show thatfeature selection is an effective process fordetecting changes basedon textural and spectral features. Each of the techniques for selecting features has its own limitations and advantages, but in general, improve the CD accuracy. The proposed GA-based feature selectionapproach was found to be relatively effective when compared withPCA and SSA approaches. Overall accuracy and Kappa coefficient ofCD were increased from 53.66% to 88.49% and 58.94% to90.39%respectivelyusing proposed methods compared tothe use of spectral information.
