When predicting landslides, it is important to understand the past landslide cases and prepare for similar case scenarios. By gathering information on frequently occurring slope disaster cases, researchers have determined that landslide behaviors change according to different conditions, including geological and topographical features, rainfall characteristics, and landslide progression (Mahr 1977, Chigira 2020). However, no system has been proposed in which these various conditions can be discussed in a unified manner. In contrast, remote sensing technology is evolving, and the surface profile obtained through light-induced direction and ranging (LiDAR) is numerical information with a history of topographic features and deformations (Demurtas et al. 2021). However, the interpretation by geomorphological and geological experts alone is not versatile enough, and a wide area cannot be determined at a time, leading to oversight. Hence, in this study, we aimed to determine the shared similar topographic features from LiDAR before the landslide.
The topographic feature related to slope failures is gravity deformation. This means continuous deformation of the bedrock occurs near the surface as well as in deep underground. Gravity deformation has been reported in the past in glacial landforms, such as zagging and double mountain ridges, in the Alps (Zischinsky, 1966). Dramis and Sorisso-Valvo (1994) and Agliardi et al. (2001) have reported deep-seated gravitational slope deformations (DSGSDs) where they defined large masses as those in which gravitational deformation is caused by small displacements with or without a slip surface. Chigira et al. (2003) showed that the 1999 Tsaoling landslide was not a reactivation of an old landslide but it had already moved slightly before the 1999 event, providing precursor evidence. Crosta et al. (2006) numerically predicted that the upper slopes of large landslides are unstable and the potential instability factors, such as landslide sliding and mobility, in the Italian Pre-Alps. While the close relationship between topographic features and DSGSDs is being clarified, a large-scale rockslide-debris avalanche that occurred in 2006 in southern Leyte in Philippines has attracted attention as a large-scale collapse. Evans et al. (2007) pointed out that the avalanche was related to the Philippine Fault and that non-brittle deformation of the surface rock was observed as a result of slope movement. Guthri et al. (2009) reported that the rockslide-debris avalanche was caused by not only heavy rainfall and a simultaneous earthquake, but also the result of progressive rupture and structural weakening. In the Kii Peninsula, which is the target area of this study, landslides have been frequently observed in recent disasters. Regarding Typhoon Talas in 2011, approximately a few million m3 of massive landmass moved in rapidly. DSGSDs were determined to be their precursors in most of the landslides (Chigira et al. 2013). Hiraishi and Chigira (2009) and Tsou et al. (2017) noted that large-scale knickpoint formation and gentle slopes influence DSGSD formation. Additionally, knickpoints form during rapid downward erosion. DSGSDs and landslides were categorized into two groups—those entirely within the paleo-surface and those across knickpoints—with 75% of the 2011 landslides being associated with the inner slopes of the valley, two to three times more than those that are entirely within the paleo-surface. Thus, the landslide has some shared topographic features due to DSGSDs. Hence, to predict potential landslide locations, it is necessary to understand DSGSD formation and determine their progression manner.
For making landslide predictions by focusing on topographical features, such as DSGSDs, landslide inventory maps (LIMs) and landslide susceptibility maps (LSMs, Can et al. 2019) can be used. High-resolution satellite images and photographs taken by unmanned aerial vehicles (UAVs), including terrestrial and UAV-mounted LiDAR (light detection and ranging) data, have become available in recent years (Ghorbanzadeh et al. 2019a). These data can easily be arranged chronologically, and the geographic data of the same location at different times can be used for analysis. Much of the geographic data are complex, it can be difficult to analyze such data at large scales, and some data may be inadvertently overlooked. Machine learning has been a method for complementing multivariate analyses and has been used in LIMs and LSMs using gridded data. LSM-based landslide location prediction has been conducted by Lee (2004), who used the sigmoid function. Dahal et al. (2008) used weights-of-evidence modeling conducted using eight elements to predict landslides. Conversely, Chen et al. (2018) reported creating an LSM using an entropy model and support vector machine (SVM).
In deep learning analyses, considering that the environmental factors about landslide susceptibility are non-correlated or have non-linear correlations, 27 environmental factors were analyzed using a fully connected spare auto-encoder for landslide susceptibility prediction (LSP) (Huang et al. 2020). Liu and Wu (2016) indicated that the Deep Auto-encoder network model method outperforms ANN and SVM in terms of precision, recall, and accuracy by evaluating the remote sensing images downloaded from Google Earth.
Images have also been used in analyses that use a convolutional neural network (CNN) instead of gridded data. Many images, as well as the engineer’s perspective, were used as datasets (Wang et al. 2019; Wang et al. 2021). Ghorbanzadeh et al. (2021) achieved an accuracy rating of almost 90% by testing the potential of CNN for slope collapse detection using Sentinel‑2 multispectral imagery. Additionally, CNN2D has the ability to outperform traditional methods such as CNN1 and SVM owing to its high prediction rate and generalizability (Youssef et al. 2022). Deep learning algorithms can prevent local optimization and eliminate the need to adjust model parameters based on autonomous processes (Carrio et al. 2017). Meanwhile, attempts have been made to search for the optimal model for the network structure of CNNs using automatically constructed high-performing architectures. This method proposes new structures that can be viewed without human coordination (Real et al. 2020). The results of this method not only reduce and streamline the human design burden, but also indicate the possibility of finding machine learning algorithms that do not use neural networks. These approaches have been used in a variety of fields, in some cases, with better results than human experts (Suganuma et al. 2020).
The purpose of this study is to improve the overall performance of the CNN2D-based model, which has been studied well in landslide susceptibility modeling (LSM), as an automatically constructed model and to systematically compare it with conventional CNN models. Furthermore, this is the first study that focusses on pre-disaster DEM images. We also compared the effectiveness of these self-search models in detail. Specifically, we evaluated the models in a test domain using a set of indicators such as receiver operating characteristic (ROC) and accuracy. We believe that the results of this study can aid researchers and local decision makers in spatial information and erosion control in reducing the local landslide risk.