It happens in a blink of an eye. One second, the ground is there; the next second it’s gone. That is the speed of a landslide: sudden, swiftand perhaps most unsettling, it strikes without warning. The UN Office for Disaster Risk Reduction ranks these natural hazards as the fifth most frequent and the seventh most damaging.
A 3.915Mb PDF of this article as it appeared in the magazin is complete with imagesis available by clicking HERE
“Similar to earthquakes, landslides are next to impossible to predict,” says Daniel Hlbling, a research scientist at the University of Salzburg’s department of geoinformatics. “And they cause significant damage. They can wipe out entire villages in a few seconds. After the event, it can be difficult to rapidly assess and map the extent of the landslide as well as find adequate tools to identify high-risk areas and create planning strategies. There are still a number of issues to resolve with landslides. They’re a puzzleand this makes them intriguing to study.”
To be sure, the puzzle of landslideshow to adequately define them, categorize them, detect them, map them and plan for themhas been an intriguing focus of much research since the late 1990s. But with the advancement of geospatial tools, such as very high-resolution satellite imagery and synthetic aperture radar interferometry (InSAR), coupled with powerful object-based image analysis (OBIA) technology, the interest in developing more effective solutions for landslide detection, mapping, inventorying, monitoring, and possibly, forecasting, has grown considerably in the last five years.
Hlbling, for one, has had an almost laser-like focus on landslide-application research since 2009. Using a host of sources like optical imagery and SAR data in combination with Trimble’s eCognition software, he has been both developing and testing automated approaches to landslide detection, mapping, inventorying and monitoringtasks that are dominated by manual, traditional methodsto determine the feasibility of using these methods operationally.
And there is considerable promise. Based on his and other colleagues’ research, Hlbling feels a groundswell of possibilities is afoot to help organizations better assess, map, prepare and plan for the unpredictable nature of landslides.
Complex by nature
The complexities in efficiently and accurately identifying, mapping and inventorying landslides are many. This is predominantly because the unpredictability of landslides means they don’t have uniform behaviors or patterns; they don’t always look and act the same. And identifying and mapping these disasters is a very individual approachwhat is or is not a landslide is decided by the expert mapping the event.
“The quality of landslide mapping is based on the mapping experttheir skills, their geographical knowledge of the area, the size of the area and the data they use,” says Hlbling. “Given the different types of landslides, their variability in shape and size, and the difficultly in distinguishing manmade features such as small quarries or harvested forests from clearings made from a landslide, the analysis and mapping can be quite labor intensive, subjective and highly inconsistent.”
However, the rise of more extreme and damaging landslide events in the last decade has also given rise to more interest in funding research that’s focused on trying to bring better efficacy, accuracy, and possibly, predictability to mapping and monitoring landslides.
Despite having no prior experience with studying landslide phenomena, Hlbling has focused on little else since beginning his research eight years ago. To date, he has conducted research in Taiwan, Italy, Austria, New Zealand and Iceland to develop new, semi-automated methods for a host of applications including classifying, inventorying and mapping landslides, detecting landslide hotspots and mapping landslide change detection. At the core of all his research has been Trimble eCognition OBIA technology.
“OBIA, with remote sensing data, is the most powerful tool for detecting and analyzing landslides,” says Hlbling. “You can integrate spectral, spatial, morphological, textual, and contextual properties in one interlinked framework. All that diverse data enables the software to mimic how the human brain identifies and categorizes objects, making it far superior to traditional pixel-based approaches which can’t do that.”
But of all the research Hlbling has conducted, there are two areas he says hold particular promise: landslide hotspot mapping and combining optical and InSAR data to better map, track and possibly predict future landslides.
Hotspot Mapping
With its combination of steep slopes, erodible hill country, frequent earthquakes, intense rainstorms, deforestation and farming-induced clear cutting, New Zealand has not only been susceptible to extensive landslide erosion, it has seen landslide activity increase by about seven times its natural rate.
The country’s local governments have diligently worked to build detailed landslide inventories and maps showing the location, extent and severity of each landslide event in order to develop effective mitigation measures. However, they have been conducting this work by traditional, manual meansvisually interpreting each aerial or satellite image and manually delineating and mapping each landslide identifiedwhich has been tedious, slow and subjective.
In 2016, Hlbling partnered with New Zealand’s Landcare Research, a Crown Research Institute headquartered in Lincoln, to test an OBIA, semi-automated approach for identifying and mapping landslide-prone “hotspots” based on historical and recent aerial photography.
The team selected an approximately 1,000-hectare (2,470-acre) study area located about five kilometers southeast of the small town of Pahiatua, in New Zealand’s North Island, a pastoral hilly region where rain-triggered landslides are common.
They acquired three black and white orthophotos, one from 1944, one from 1979 and one from 1997 as well as two natural color orthophotos from 2005 and 2011, which had nominal accuracies of 15 m and a spatial resolution of up to 0.4 m. Additionally, they obtained a 15-m-resolution DEM to provide ancillary data, such as slope information.
In order to adequately compare the OBIA approach to the manual approach, Landcare researchers spent two weeks manually digitizing visible landslides on each orthophoto in ArcGIS.
In parallel, Hlbling spent one day preparing Trimble’s eCognition software for integrating the datasets and classifying the landslides. Since aerial photographs from five different points in time were used, Hlbling developed a single mapping routine that could be applied to all images, using the 2011 orthophoto as the master. The customized rule set used spectral, spatial, contextual and morphological properties to properly classify and map all detected landslides on the 2011 orthophoto. Confirming that the workflow was successful, he then applied that rule setwith a few modificationsto the other orthophotos. In a few hours, eCognition classified all visible landslides across all five time stamps.
With both the manual and automated mapping completed, the team compared the two approaches and found the eCognition mapping was on par with the manual results. Given the subjective nature of manual mapping, it was difficult to conclude whether the automated mapping was more accurate, but what was clear, is that the eCognition approach is considerably faster, more consistent, more objective, and it’s easily repeatable.
“The manual mapping was painfully slow,” says Harley Betts, a researcher with the soils and landscapes team at Landcare Research’s Palmerston North office. “The eCognition approach has the potential to cut out a big chunk of that manual stage. I was very impressed with that.”
As a complement to manual methods, an automated system would both allow experts to quickly detect and inventory landslides over large areas and identify unstable areas. That could, over time, lead to predictive modelling.
“If you follow the idea that the past can be the best indicator for the future, then by studying the historical evolution of landslide hotspots and mapping the changes over time, you could use the hotspot mapping, to some degree, for prediction,” says Hlbling. “The hotspot maps could help flag landslideprone areas, giving valuable insight to land management experts to help them assess landscape dynamics, create location-specific, risk mitigation measures, and potentially, forecast where landslides might occur.”
Morphing into automation
Similar to New Zealand, Iceland is no stranger to landslides. With its volcanic, tectonic and glacial activity, combined with extreme storm events, eastern and northern Iceland are routinely on alert for possible landslides. In particular, “creeping landslides,” or deep-seated, slow-moving slides are a continual threat.
These characteristics made the country a particularly suitable study area for the MORPH (Mapping, Monitoring and Modelling the Spatio-Temporal Dynamics of Land Surface Morphology) project, one of Hlbling’s most recent studies that aims to develop an efficient and transferable OBIA method for mapping slope instabilities, including landslides, and volcanic deposits. Although the overall objective is to combine multi-sensor imagery from optical and radar satellites into an eCognition mapping workflow, unique to this endeavor is the pairing of optical satellite imagery with InSAR datasets in eCognition to create a more powerful, integrated landslide tool.
“Typically, shallow, quick-moving landslides are visible on optical imagery so they are easier to map,” says Hlbling. “But often, the land continues to slowly shift and move after an initial landslide and you can’t readily see that on optical imagery. InSAR data derived from radar satellite imagery can detect ground-surface movements of centimeters and even millimeters, which can show us where deep-seated landslides are and how quickly they’re moving. That combination could result in a more comprehensive and detailed landslide inventory.”
To test the viability of this integrated mapping approach, the team chose a site in the rfajkull region, a southeastern area that is susceptible to unstable slopes, dynamic landscape movements, soil erosion and floodsa ripe recipe for landslides.
For this initial study, the team used a 5-m-resolution optical RapidEye image, a 2-m-resolution LiDAR-derived DEM and two 3-m-resolution TerraSAR-X StripMap scenes. In addition to calculating a vegetation index from the optical image and slope values from the DEM, they used the two SAR scenes to calculate the phase difference between the two images, which helps identify areas on the ground surface that have moved.
Hlbling and his team developed an eCognition rule set to integrate the imagery and InSAR data information to identify and map all landslides as well as designate areas potentially affected by landslides. The software not only distinguished landslides based on the RapidEye image but with the additional InSAR data, it identified more potentially affected landslide areas. That approach opens up “remarkable possibilities” for improved landslide mapping, says Hlbling.
“The additional detail and movement information of InSAR could help us produce more refined landslide maps,” he says. “You can map the visible landslides with the optical imagery and the `invisible’ slow-moving slides with the SAR data and put these two datasets together to create a complete landslide inventory. And by using the velocity detail from InSAR, we can analyze land movement over time and potentially forecast how the landslide might continue to move based on the historical movement. That’s incredibly promising.”
As the MORPH work won’t conclude until 2019, the true assessment of this automated, integrated method will have to wait. But Hlbling is optimistic, both about these initial results and future ones to come.
“From the beginning, a goal has been to develop a more efficient and reliable mapping framework that we can port to other countries vulnerable to landslides,” he says. “Each study gives us the opportunity to extend our knowledge, test the capabilities of OBIA technology and refine our tools, all of which could lead to significant, positive shifts in our approaches to detecting and mapping landslides.”
That kind of movement would no doubt be welcome.
Mary Jo Wagner is a Vancouver-based freelance writer with 25 year’s experience in covering geospatial technology.
Sidebar:
It’s time for Taiwan
If there is one country more invested in finding more efficient and accurate methods for identifying, inventorying and mapping landslides, it’s Taiwan, a country that’s battered by three to five typhoons nearly every year.
Landslide specialists at Taiwan’s Disaster Prevention Research Center (DPRC), who use a combination of manual interpretation and pixel-based mapping systems, have been working with Hlbling since 2009 to develop a more automated approach to classifying and mapping landslides.
Hlbling has held several workshops illustrating OBIA technology and conducted nearly 10 studies to test the viability of an Trimble eCognition-based methodology. In one study, Hlbling produced map results in 2.5 hours for an area that took 12 hours to map manually. Hlbling says DPRC users have been impressed by the OBIA approach.
“The time aspect is a very important issue for them,” he says. “They need to quickly acquire information about new landslides after landslide triggering events for their disaster management operations.”
And research continues. Taiwan is part of a three-year study that began in August that aims to analyze a timeseries of optical satellite imagery to determine spatio-temporal hotspots of landslide-induced river-course changes.
A 3.915Mb PDF of this article as it appeared in the magazine is HERE