The mountains of Taiwan, with the highest peaks rising to almost 4000 m a.s.l., are characterized by fractured rock formations, high relief and steep stream gradients. These mountains, particularly the central mountain range (CMR), influence the tracks and intensity of typhoon events [1
]. During summer and autumn, Taiwan is regularly affected by typhoons (tropical cyclones), three to four per year on average, which bring heavy rainfall [2
]. Some studies indicate that the number of typhoons hitting Taiwan increased after the year 2000, also resulting in heavier precipitation in recent years [3
]. Heavy rainfall is the main landslide triggering factor in Taiwan [5
]. The rainfall events are usually associated with typhoons, which account for almost 50% of the island’s total rainfall [6
]. Studies have found relationships between long-duration and moderate intensity rainfall events and large and deep-seated landslides [8
], sediment yield and debris flows [5
Large, rapid landslides and debris flows frequently lead to fatalities. Major endeavors are necessary to protect people and settlements in areas at risk, and to implement prevention and early warning measures [10
]. In general, as well as in Taiwan, such catastrophic landslides also cause severe damages to infrastructure, and efforts both in time and money are needed to recover and maintain the transportation infrastructure such as roads and bridges [14
]. Beyond the direct landslide hazard, large landslides can initiate natural hazard cascades by damming rivers and inducing catastrophic flash floods and debris flows [18
]. The large amount of mobilized debris that originates from landslides significantly affects the drainage system, for example, resulting in an increase in erosion and sediment discharge in rivers, and changes in channel size and shape [5
]. According to Chen et al. [9
] about 384 Mt y−1
of sediment is transported into the ocean in Taiwan, whereby the high proportion of large landslides significantly contributes to this high annual sediment yield.
In August 2009, typhoon Morakot caused a record-breaking cumulative rainfall (more than 2000 mm in three days), which led to debris flows and mudflows, flooding in coastal areas, and massive landslides [23
]. The rainfall also triggered one of the most famous and fatal landslides in Taiwan, the Xiaolin landslide [25
]. The Butangbunasi landslide [29
] is another example of a large rainfall-triggered landslide in Taiwan. Reactivation and extension of this landslide have been resulting in repeated sediment delivery to the Laonong River, especially during torrential rainfall. The river course has been frequently affected, leading to the formation of a landslide-dammed lake several times during the past three decades [29
]. The magnitude of the Butangbunasi landslide is even significantly larger than of the disastrous Xiaolin landslide [29
]. A deeper knowledge of the evolution of landslides and their triggering factors is crucial for hazard mitigation [37
]. Therefore, mapping and analyzing the evolution of such large landslides over time helps to better understand their reactivation rates and their impact on downstream areas.
Remote sensing plays a key role in studying landslides and provides an adequate and cost-effective source to derive information about landslide distribution and types [39
]. The use of remote sensing data also helps to investigate the potential impacts of landslides such as the damming of rivers, particularly in difficult to access and remote mountain regions [43
]. The value of remote sensing for landslide studies becomes more and more evident with the increasing amount of freely available Earth observation (EO) data, which provides remarkable opportunities to map and monitor landslides over time [44
Object-based image analysis (OBIA) provides a suitable methodological framework for efficient landslide mapping, as well as landslide change analysis [47
]. By working on the object-level instead of the pixel-level, OBIA allows considering spectral, spatial, textural, morphometric and hierarchical properties for the classification of landslides [49
]. Moreover, it is argued that using OBIA yields better classification accuracies than pixel-based classifications [51
]. Several studies employed OBIA for landslide mapping and landslide change detection in Taiwan [32
], but none of them used time series of images for investigating the evolution and reactivation of an active large landslide.
The aim of this study is to analyze the evolution of the Butangbunasi landslide in south-central Taiwan using OBIA and time series of freely available Landsat images and to investigate the potential correlation between changes in landslide area and heavy rainfall events during selected typhoon (including tropical storm) events.
Semi-automated techniques can limit the subjectivity in landslide mapping and can contribute to improving the reproducibility of landslide maps [78
]. OBIA is such a technique and provides a set of suitable tools to semi-automatically map the evolution of landslides with time series of satellite images. The developed OBIA workflow was designed to be transferable across images, whereby only minor modifications for each Landsat sensor were necessary. This reduces the analysis time and increases the transferability of the approach. We used spectral indices for the landslide classification supported by ancillary DEM data to avoid the classification of specific false positives. While a DEM was available for only one point in time for our analysis, using DEMs acquired after each triggering event would increase the classification accuracy and would also allow a volume change estimation. In practice, however, multi-temporal DEM data are rarely available. The semi-automated mapping led to reasonable results. However, the determined accuracy values need to be considered with care, since any reference data created by manual mapping includes a certain degree of uncertainty and subjectivity [45
The Landsat archive offers multispectral imagery since the 1980s suitable to identify recurrent changes in the area of large landslides such as the Butangbunasi landslide. Even if the spatial resolution of 30 m does not allow to identify very small changes, trends over time can be well depicted. However, the exact timing of landslide extension/reactivation following typhoons or tropical storms remains difficult. We employed the first cloud-free Landsat image acquired after such an event, but in some cases, the time span between a rainfall event and image acquisition date was up to several months or even longer. The accuracy of the OBIA mapping probably also depends on the time elapsed between a landslide triggering event and the acquisition of the next satellite image [37
]. A short delay would allow deriving more detailed information about the landslide reactivating, the revegetation time and the formation of landslide-dammed lakes. The time span of over two years between the last image used and the last identified typhoon is probably the reason for the slightly decreasing trend in landslide area since initial revegetation happens quite fast. Several studies investigated the vegetation recovery after the occurrence of landslides. For example, Lin et al. [79
] estimated a vegetation recovery rate of approximately 60% two years after landslides and Chou et al. [80
] found a vegetation recovery rate of approximately 90% six years after landslides in central Taiwan. The new generations of satellites, for example, freely available data such as Sentinel-2 or EO data from commercial data providers, already provide a higher temporal and spatial resolution for recent years. This offers great opportunities for improving studies similar to the presented one in the future when longer time series of very high resolution (VHR) images will be available.
Rainfall-triggered landslides are particularly frequent in areas heavily affected by typhoon events, such as Taiwan. Freely available rainfall data are essential to improve early-warning models, as new technologies and research opportunities emerge. In this study, we focused on two free and open rainfall data sources. The CHIRPS data are globally available since 1981, however, its coarse spatial and temporal resolution limits its usage for local studies. In addition, the CHIRPS data under/overestimate daily precipitation. In our case, rain gauge data available from the CWB typhoon database were a more suitable alternative. Our results indicate that the duration of the heavy rainfall event is the main parameter linked to the landslide area change, while cumulative rainfall and mean intensity did not show significant correlations with the extension of the Butangbunasi landslide. Further analyses should be performed to find a direct causation between the rainfall event duration and the landslide area change. However, the freely available CWB data are limited to those hours when the typhoon passed Taiwan, and hence a complete historical rainfall database cannot be analyzed together with these extreme events. Several parameters could be computed from such a database, as indicated by Guzzetti et al. [81
], which could lead to a more robust evaluation of the relationship between rain events and landslide evolution.
Better knowledge about the reactivation of large landslides and the recurring impact on downstream areas is of high importance for disaster mitigation. Even if our results did not indicate a direct relationship between the extension of the Butangbunasi landslide and the strength of the typhoon event, it became evident that also comparatively small typhoons or tropical storms could cause landslide reactivation. This is, for example, of high relevance for implementing early warning measures. The repeated sediment delivery after rainfall events frequently impacts the rivers system, which can result in the formation of landslide-dammed lakes and debris flows, and eventually poses a risk for people, settlements and infrastructure downstream. Major efforts are taken to maintain the transportation infrastructure in the study area and to avoid the repeated damming of the river (Figure 8
). In particular, the Southern Cross-Island Highway, which is a popular tourist route that crosses the CMR and that provides a connection to the remote Yushan National Park, was severely affected by typhoon Morakot and following events, and the associated debris and sediment from the Butangbunasi landslide [29
]. Information about the evolution of the Butangbunasi landslide is thus also important for planning and implementing maintenance and reconstruction activities of roads and bridges.
In this study, we focused on one large landslide. By studying larger areas and relating spatio-temporal landslide hotspots to rainfall events brought by typhoons or tropical storms, more robust correlations between landslide extensions and triggering events might be found. Further research should also emphasize on the combination of OBIA and machine learning approaches for automated landslide time series analysis.
Large rainfall-induced landslides are among the most dangerous natural hazards in Taiwan, putting people and infrastructure at risk. Thus, better knowledge about the evolution of large landslides, their triggering factors and their potential to initiate cascading hazards is important in several respects.
Often insufficient information exists on landslide occurrence and reactivation intervals. Findings from the analysis of time series of satellite imagery, as provided for example by the Landsat missions, can provide useful information for supporting hazard mitigation and spatial planning. At the same time, the constantly increasing amount of satellite imagery at higher spatial and temporal resolutions implies the need for efficient and robust landslide (change) mapping methods. OBIA provides a suitable methodological framework for addressing these challenges. In this study, we semi-automatically mapped the evolution of the Butangbunasi landslide using Landsat time series data. The OBIA mapping results showed that the most significant landslide extension happened after typhoon Morakot in 2009. Freely available rainfall data were analyzed to find potential relationships between the reactivation and evolution of the landslide area and heavy rainfall during typhoon events. Our results indicate that the duration of the heavy rainfall event is the main parameter linked to the landslide area change, whereas daily precipitation, cumulative rainfall and mean intensity did not present strong significant correlations.
While landslides and associated hazards are a significant problem under present-day climate regimes, it is likely that climate change will lead to more frequent and extreme landslide-triggering events such as typhoons and tropical storms in Taiwan. Consequently, even more landslides may occur in the future. With this in mind, the relevance of studies that investigate and analyze the evolution of large landslides with respect to triggering events becomes even more important. Respective results can serve as input for hazard and risk analysis and the implementation of prevention and mitigation measures.