Abstract:
Objectives The 2008 Wenchuan Ms 8.0 earthquake triggered tens of thousands of landslides and produced about 10 billion cubic meters of loose material, resulted in a significant increase in the scale and frequency of landslides after the earthquake. The temporal and spatial evolution of post-earthquake landsliding has become the focus of attention of scholars and the public. The Mianyuan River basin is selected as a case study, and the temporal and spatial evolution trend of post-earthquake landsliding is quantitatively analyzed using multi-temporal satellite images.
Methods Firstly, the satellite images of Landsat 7, SPOT-5, RapidEye and Planet from 2007 to 2018 are collected and preprocessed such as atmospheric correction, ortho-rectification and image cutting, etc. Then, the maximum likelihood method and the random forest algorithm are engaged to automatically detect the landsliding in the Mianyuan River basin using the preprocessed satellite images. Finally, the landslide detection accuracy of the two methods is evaluated by comparing the landslides interpreted manually and the landslides detected automatically.
Results The average recognition accuracy of random forest algorithm is 86.73%, while the average recognition accuracy of maximum likelihood method is 73.11%. The total area of new landslides detected by the random forest algorithm in 2007, 2008, 2011, 2013, 2015, 2016, 2017 and 2018 are 0.24, 51.54, 19.25, 7.21, 6.39, 7.35, 3.51, 3.82 km2, respectively.
Conclusions The random forest algorithm method has higher detection accuracy than the maximum likelihood method. The landsliding activity in the study area during the 10 years after the Wenchuan earthquake can be divided into three stages: 2008 to 2011 is the strong activity period, the scale and frequency of landsliding activity are significantly increased than before the earthquake, and the disaster losses are serious; 2012 to 2016 is the moderate activity period, the scale and frequency of landsliding activity gradually decreased, and disaster events occurred occasionally; 2017 to 2018 is the weak activity period, the scale and frequency of landsliding activity decreased significantly. After 2017, disaster events rarely occurred in the study area, but landsliding activity still did not return to the pre-earthquake level.