Ahmad Malek Nejad Yazdi; Hassan Ghassemian; Vahid Esavi
Volume 7, Issue 3 , November 2015, , Pages 65-82
Abstract
Most of common classification algorithms in remote sensing are based on spectral characteristics of the pixels. These approaches result in ignorance of many precious information, such as texture, in the classification process. The urban environment has an inhomogeneous texture, which makes land covers ...
Read More
Most of common classification algorithms in remote sensing are based on spectral characteristics of the pixels. These approaches result in ignorance of many precious information, such as texture, in the classification process. The urban environment has an inhomogeneous texture, which makes land covers detection a complicated process. In this study, use of texture extracted from the panchromatic image of ALI detector for improvement of Hyperion image's classification accuracy in urban regions was analyzed. Classification carried out using Random Forests method and in five different scenarios. These scenarios included: 1- Classification of the fused image by CNT method (Without Incluion of Texture Information), The other four scenarios covered the classification used by simultaneous use of texture extracted by Gray Level Co-occurrence Matrix »GLCM« in 4 different window sizes: 3,5,7,9 and fused image. Results of these analyses revealed that use of texture information as a useful parameter can lead to an enormous improvement in classification accuracy. Our findings showed that use of texture resulted in an increase in overall accuracy by around 10 percent from 80.47 to 90.74 percent . Many of land use/land covers such as roads, residential and industrial areas also experienced the improvement in producer and user accuracies. OOB error as an essential random forests parameter inclined as far as 11 percent from 19.86 to 8.87 percent. Moreover, the GLCM feature vectors such as mean and contrast achieved high ranks in importance evaluation in random forests classification. Increase of window size also led to a rise of classification accuracy and the window size 9 gained the highest accuracy accordingly.
