Semi-Automatic Landslide Detection using Sentinel-2 Imagery: Case Study in the Añasco River Watershed, Puerto Rico
Abstract
The increasing availability and spatiotemporal resolution of satellite imagery has propelled efforts to use this imagery to automatically detect landslides. Many of the proposed workflows require specialized software making them inaccessible to many geoscientists. Additionally, the outcome of competing approaches is challenging to compare. The goal of our study is to use accessible software to both detect landslides using satellite imagery and evaluate our workflow using robust evaluation metrics.
Hurricane Maria devastated islands in the Caribbean in 2017 and triggered 70,000 landslides across the island of Puerto Rico. After the hurricane, the SLIDES-PR group at the University of Puerto Rico Mayagüez manually mapped landslides throughout Puerto Rico using satellite imagery. Using their data as a training set, we focused on a 46 km2 training site within the Añasco River Watershed in Puerto Rico to develop our methods and then applied them across the 415 km2 watershed. Inputs include 10 m ESA Sentinel-2 satellite imagery and digital elevation data. Our workflow detected landslides triggered by the hurricane using the change-detection and machine-learning classification methods in ESRI ArcGIS 10.7 and Google Earth Engine. The results illustrate that Sentinel-2 data can be used to automatically detect landslides but the software we used needs improved image segmentation and thresholding methods to produce more accurate results. While most studies use similar accuracy metrics, such as recall and precision, we found that there is a lack of consistency in how these measures are used. Inconsistency in the way these measures are quantified makes it difficult to compare the accuracy of different methods among studies. In this study we propose exploring accuracy metrics that use the intersection over union (IOU) metric, commonly used in the field of computer vision, to determine the accuracy of individual landslides in regard to recognition and extent. To summarize, efforts presented here highlight the need for a shift toward more open and accessible means of landslide detection for the broader geoscience community to fully recognize the benefits of automatic and semiautomatic landslide detection.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMNH14A..06M
- Keywords:
-
- 9805 Instruments useful in three or more fields;
- GENERAL OR MISCELLANEOUS;
- 1920 Emerging informatics technologies;
- INFORMATICS;
- 4314 Mathematical and computer modeling;
- NATURAL HAZARDS;
- 4333 Disaster risk analysis and assessment;
- NATURAL HAZARDS