Sensitivity analysis of automatic landslide mapping: numerical experiments towards the best solution

Abstract

The automatic detection of landslides after major events is a crucial issue for public agencies to support disaster response. Pixel-based approaches (PBAs) are widely used in the literature for various applications. However, the accuracy of PBAs in the case of automatic landslide mapping (ALM) is affected by several issues. In this study, we investigated the sensitivity of ALM using PBA through digital terrain models (DTMs). The analysis, carried out in a study area of Poland, consisted of the following steps: (1) testing the influence of selected DTM resolutions for ALM, (2) assessing the relevance of diverse landslide morphological indicators for ALM, and (3) assessing the sensitivity to landslide features for a selected size of moving window (kernel) calculations for ALM. Ultimately, we assessed the performance of three classification methods: maximum likelihood (ML), feed-forward neural network (FFNN), and support vector machine (SVM). This broad analysis, as combination of grid cell resolution, surface derivatives calculation, and performance classification methods, is the challenging aspect of the research. The results of almost 500 experimental tests provide valuable guidelines for experts performing ALM. The most important findings indicate that feature sensitivity in the case of kernel size increases with coarser DTM resolution; however, the peak of the optimal feature performance for the selected study area and landslide type was demonstrated for a resolution of 20 m. Another finding indicated that in combining a set of topographic variables, the optimal performance was acquired for a DTM resolution of 30 m and the support vector machine classification. Moreover, the best performance of the identification is represented for SVM classification.

Geographic Object-Based Image Analysis for Automated Landslide Detection Using Open Source GIS Software

“…We successfully applied and assessed an automated open source landslide detection procedure solely based on HRDTM-derived data. The moderate classification agreement based on the pixel-level assessment (κ of 0.42, and 0.48 in post-processed classification) is consistent with previous research using similar methods [13,15] (κ between 0.45 and 0.6). Although different performance measures and data sets were used in studies within commercial software, they also showed a moderate predictive skill [7,18] (percentage of inventoried landslides detected correctly at the object level 69%, after post-processing 66%; 71% in [7] and 68.7% and 73.3% in [18]).…”

“…For landslide classification we used the support vector machine, a flexible supervised machine-learning technique [56]. SVM has become a widely used classification method that has shown its potential in landslide distribution modeling [7,13,21]. We used the e1071 package in R [57] within the mlr modeling framework.…”

“…Following Richter [59], cost was drawn between 2 −12 and 2 15 , and gamma between 2 −15 and 2 6 . As performance criterion we chose Cohen's Kappa (κ)-a well-established accuracy measure for classification problems [60], which was also applied in similar studies [13,15].…”

“…Leshchinsky et al [11] and Bunn et al [12] developed and extended the contour connection method, a vector-based method to detect different landslide features using a mesh of nodes and connecting lines. Pawłuszek et al [13] detected landslides using a pixel-based classification scheme, and benchmarked different classifiers while also accounting for different grid resolutions. However, with high-resolution data, pixel-based classification approaches are often confronted with noise inherent in the data ("salt-and-pepper effect"), resulting in a high false positive rate.…”

“…Moreover, previous landslide detection procedures were mostly implemented in commercial software environments such as Definiens eCognition [7,15,16,18,21], ESRI ArcGIS [13,15], the ENVI image analysis software [13,15], or MathWorks MATLAB [9,21]. Only few authors made their algorithms publicly available [9,18], or used computing environments released under open source licenses [11].…”