Landslide Trail Extraction Using Fire Extinguishing Model

Landslide trails are important elements of landslide inventory maps, providing valuable information for landslide risk and hazard assessment. Compared with traditional manual mapping, skeletonization methods offer a more cost-efficient way to map landslide trails, by automatically generating centerlines from landslide polygons. However, a challenge to existing skeletonization methods is that expert knowledge and manual intervention are required to obtain a branchless skeleton, which limits the applicability of these methods. To address this problem, a new workflow for landslide trail extraction (LTE) is proposed in this study. To avoid generating redundant branches and to improve the degree of automation, two endpoints, i.e., the crown point and the toe point, of the trail were determined first, with reference to the digital elevation model. Thus, a fire extinguishing model (FEM) is proposed to generate skeletons without redundant branches. Finally, the effectiveness of the proposed method is verified, by extracting landslide trails from landslide polygons of various shapes and sizes, in two study areas. Experimental results show that, compared with the traditional grassfire model-based skeletonization method, the proposed FEM is capable of obtaining landslide trails without spurious branches. More importantly, compared with the baseline method in our previous work, the proposed LTE workflow can avoid problems including incompleteness, low centrality, and direction errors. This method requires no parameter tuning and yields excellent performance, and is thus highly valuable for practical landslide mapping.

Distribution of landslides in the area of the town of Polski Trambesh, Northern Bulgaria

The article presents the results of the landslide mapping in the area of the town of Polski Trambesh, central Northern Bulgaria. A total of 37 new landslides have been described, which have not been included in the landslide register so far. Landslides are classified by area, type and activity. Their location is mapped in a GIS environment.

The Relevance of Geotechnical-Unit Characterization for Landslide-Susceptibility Mapping with SHALSTAB

Given the increasing occurrence of landslides worldwide, the improvement of predictive models for landslide mapping is needed. Despite the influence of geotechnical parameters on SHALSTAB model outputs, there is a lack of research on models’ performance when considering different variables. In particular, the role of geotechnical units (i.e., areas with common soil and lithology) is understudied. Indeed, the original SHALSTAB model considers that the whole basin has homogeneous soil. This can lead to the under-or-overestimation of landslide hazards. Therefore, in this study, we aimed to investigate the advantages of incorporating geotechnical units as a variable in contrast to the original model. By using locally sampled geotechnical data, 13 slope-instability scenarios were simulated for the Jaguar creek basin, Brazil. This allowed us to verify the sensitivity of the model to different input variables and assumptions. To evaluate the model performance, we used the Success Index, Error Index, ROC curve, and a new performance index: the Detective Performance Index of Unstable Areas. The best model performance was obtained in the scenario with discretized geotechnical units’ values and the largest sample size. Results indicate the importance of properly characterizing the geotechnical units when using SHALSTAB. Hence, future applications should consider this to improve models’ predictivity.

Beyond 2D landslide inventories and their rollover: synoptic 3D inventories and volume from repeat lidar data

Efficient and robust landslide mapping and volume estimation is essential to rapidly infer landslide spatial distribution, to quantify the role of triggering events on landscape changes, and to assess direct and secondary landslide-related geomorphic hazards. Many efforts have been made to develop landslide mapping methods, based on 2D satellite or aerial images, and to constrain the empirical volume–area (V–A) relationship which, in turn, would allow for the provision of indirect estimates of landslide volume. Despite these efforts, major issues remain, including the uncertainty in the V–A scaling, landslide amalgamation and the underdetection of landslides. To address these issues, we propose a new semiautomatic 3D point cloud differencing method to detect geomorphic changes, filter out false landslide detections due to lidar elevation errors, obtain robust landslide inventories with an uncertainty metric, and directly measure the volume and geometric properties of landslides. This method is based on the multiscale model-to-model cloud comparison (M3C2) algorithm and was applied to a multitemporal airborne lidar dataset of the Kaikōura region, New Zealand, following the Mw 7.8 earthquake of 14 November 2016. In a 5 km2 area, the 3D point cloud differencing method detects 1118 potential sources. Manual labeling of 739 potential sources shows the prevalence of false detections in forest-free areas (24.4 %), due to spatially correlated elevation errors, and in forested areas (80 %), related to ground classification errors in the pre-earthquake (pre-EQ) dataset. Combining the distance to the closest deposit and signal-to-noise ratio metrics, the filtering step of our workflow reduces the prevalence of false source detections to below 1 % in terms of total area and volume of the labeled inventory. The final predicted inventory contains 433 landslide sources and 399 deposits with a lower limit of detection size of 20 m2 and a total volume of 724 297 ± 141 087 m3 for sources and 954 029 ± 159 188 m3 for deposits. Geometric properties of the 3D source inventory, including the V–A relationship, are consistent with previous results, except for the lack of the classically observed rollover of the distribution of source area. A manually mapped 2D inventory from aerial image comparison has a better lower limit of detection (6 m2) but only identifies 258 landslide scars, exhibits a rollover in the distribution of source area of around 20 m2, and underestimates the total area and volume of 3D-detected sources by 72 % and 58 %, respectively. Detection and delimitation errors in the 2D inventory occur in areas with limited texture change (bare-rock surfaces, forests) and at the transition between sources and deposits that the 3D method accurately captures. Large rotational/translational landslides and retrogressive scars can be detected using the 3D method irrespective of area's vegetation cover, but they are missed in the 2D inventory owing to the dominant vertical topographic change. The 3D inventory misses shallow (< 0.4 m depth) landslides detected using the 2D method, corresponding to 10 % of the total area and 2 % of the total volume of the 3D inventory. Our data show a systematic size-dependent underdetection in the 2D inventory below 200 m2 that may explain all or part of the rollover observed in the 2D landslide source area distribution. While the 3D segmentation of complex clustered landslide sources remains challenging, we demonstrate that 3D point cloud differencing offers a greater detection sensitivity to small changes than a classical difference of digital elevation models (DEMs). Our results underline the vast potential of 3D-derived inventories to exhaustively and objectively quantify the impact of extreme events on topographic change in regions prone to landsliding, to detect a variety of hillslope mass movements that cannot be captured by 2D landslide mapping, and to explore the scaling properties of landslides in new ways.

Evaluation of Machine Learning Algorithms for Object-Based Mapping of Landslide Zones Using UAV Data

Landslides are a critical geological phenomenon with devastating and catastrophic consequences. With the recent advancements in the geoinformation domain, landslide documentation and inventorization can be achieved with automated workflows using aerial platforms such as unmanned aerial vehicles (UAVs). As a result, ultra-high-resolution datasets are available for analysis at low operational costs. In this study, different segmentation and classification approaches were utilized for object-based landslide mapping. An integrated object-based image analysis (OBIA) workflow is presented incorporating orthophotomosaics and digital surface models (DSMs) with expert-based and machine learning (ML) algorithms. For segmentation, trial and error tests and the Estimation of Scale Parameter 2 (ESP 2) tool were implemented for the evaluation of different scale parameters. For classification, machine learning algorithms (K- Nearest Neighbor, Decision Tree, and Random Forest) were assessed with the inclusion of spectral, spatial, and contextual characteristics. For the ML classification of landslide zones, 60% of the reference segments have been used for training and 40% for validation of the models. The quality metrics of Precision, Recall, and F1 were implemented to evaluate the models’ performance under the different segmentation configurations. Results highlight higher performances for landslide mapping when DSM information was integrated. Hence, the configuration of spectral and DSM layers with the RF classifier resulted in the highest classification agreement with an F1 value of 0.85.

Active-Learning Approaches for Landslide Mapping Using Support Vector Machines

Ex post landslide mapping for emergency response and ex ante landslide susceptibility modelling for hazard mitigation are two important application scenarios that require the development of accurate, yet cost-effective spatial landslide models. However, the manual labelling of instances for training machine learning models is time-consuming given the data requirements of flexible data-driven algorithms and the small percentage of area covered by landslides. Active learning aims to reduce labelling costs by selecting more informative instances. In this study, two common active-learning strategies, uncertainty sampling and query by committee, are combined with the support vector machine (SVM), a state-of-the-art machine-learning technique, in a landslide mapping case study in order to assess their possible benefits compared to simple random sampling of training locations. By selecting more “informative” instances, the SVMs with active learning based on uncertainty sampling outperformed both random sampling and query-by-committee strategies when considering mean AUROC (area under the receiver operating characteristic curve) as performance measure. Uncertainty sampling also produced more stable performances with a smaller AUROC standard deviation across repetitions. In conclusion, under limited data conditions, uncertainty sampling reduces the amount of expert time needed by selecting more informative instances for SVM training. We therefore recommend incorporating active learning with uncertainty sampling into interactive landslide modelling workflows, especially in emergency response settings, but also in landslide susceptibility modelling.