A study on a method for detecting landslide areas automatically using constellation satellite optical images

Kazuya FUNAKOSHI, Satomi KAKUTA, Wakana YAGI, Masashi TAI,
Masayuki MATSUDA, Satoko HATTORI, Kazuo ODA and Jonghwan KIM


In this study, we tested a method to detect landslide areas automatically by using pre-and post-disaster optical satellite images. Asakura City and Toho Village in Fukuoka Prefecture as well as Hita City in Oita Prefecture were selected as test sites since many landslide disasters occurred after heavy rainfall in Northern Kyushu District in July 2017. The method utilizes a constellation of satellites equipped with high resolution optical sensors, which acquired high frequency pre-and post-disaster images. For early detection of landslide disasters, a change detection method (detect landslide area by using the change of the NDVI value between pre-and post-disaster) and a deep learning detection method (detect landslide areas using post-disaster images) was examined. And from the test sites analysis, the logical product (AND condition) which combines the two methods was the most suitable. To verify the validity, we applied the method to Iburi eastern district in Hokkaido where many landslide occurred by a large earthquake. Results show that the user accuracy was 68.1%, while the producer accuracy was 69.7%. Furthermore, analysis revealed that the detection error rate based on the land use classification was low in hillslopes covered with vegetation. However, the detection error rate was high in cultivated areas with exposed soil after harvest, and water areas with high turbidity. The study results indicate that the landslide detection method, and high frequency pre-and post-disaster optical satellite images are effective for early detection of landslide disasters that occur in wide areas at hillslopes covered with vegetation.

Key words

disaster monitoring, micro-satellite constellation, change detection and deep learning, conditional GAN