Document Type : Original Article

Authors

• marzieh nikjoui ¹
• Ali Najafinejad ²
• Ali Shamsoddini ³
• Hamid Reza Pourghasemi ⁴
• Chooghi Bairam Komaki ⁵

¹ PhD student in Watershed Management, Department of Watershed & Rangeland Management, Gorgan University of Agricultural Sciences & Natural Resources, Iran.

² Gorgan University of Agricultural Sciences and Natural Resources

³ Associate Prof. of Remote Sensing and Geographic Information System, Department of Humanities, Tarbiat Modares University, Iran.

⁴ 4. Prof. of Watershed Science and Engineering, Department of Agriculture, Shiraz University, Iran.

⁵ Assistant Prof. of Desert Department, Department of Pasture and Watershed Management, Gorgan University of Agricultural Sciences and Natural Resources, Iran.

Abstract

This research aims to develop a semi-automated method based on remote sensing and data mining algorithms on improved images in order to identify landslides in Mohammad Abad forest watershed in Golestan province. This area is prone to landslides due to the special conditions of topography and human manipulations, and every year it causes significant financial losses to residents and infrastructure destruction. In this study, two Gaofen-1 satellite images from June 1402 and March 1401 were used. Due to the different imaging seasons, all the processes were done separately on two images and finally combined for the entire study area. In order to increase the clarity of the images, three image combining methods named Brovey, PCA and Wavelet-PCA were tested. The Wavelet-PCA method was selected as the best method of image synthesis with a correlation coefficient of 97% and the closest entropy value to the original image. In the next stage, 218 landslide cases were recorded in the region through field visits, 70% of which were used for model training and 30% for model validation. Image segmentation was done along with optimization of scale parameter by local variance method and shape and compression parameters by trial and error. The optimal parameters included scale = 33, shape = 0.6 and compression = 0.5. Then the selection of features was done from textural, spectral, height, geometric and auxiliary layers using the random forest method and 16 main features were selected from among the 53 extracted features. Finally, the selected images were

Keywords

• landslide
• base object
• data mining
• scale parameter
• fusion

Main Subjects

• Remote sensing and Gis