Landslides are naturally occurring events that can lead to significant destruction and loss. They are often exacerbated by human activities such as deforestation, construction, and poor land management practices, which destabilize the earth’s surface and increase the risk of landslides. To mitigate the impact of such events, it is crucial to implement effective measures to reduce and manage risk.
Numerous studies have been conducted worldwide, employing various methodologies to assess and predict landslide susceptibility. These techniques include statistical approaches, which analyze historical landslide data to identify patterns and correlations; deterministic approaches, which use physical models to simulate landslide mechanisms; and geomorphological analysis, which examines landforms and geological processes that contribute to landslide occurrence.
Despite their usefulness, these methods have certain limitations. Challenges such as insufficient knowledge about the geological and environmental characteristics of specific areas of interest can hinder accurate assessments. Additionally, the subjectivity involved in weighing various environmental variables may lead to inconsistent results. Many traditional methods may also be more suited to simpler types of
landslides, limiting their broader applicability. Furthermore, issues regarding the reproducibility of results can affect the reliability of landslide susceptibility maps generated through these approaches.
To address these limitations, researchers have begun integrating advanced computational techniques in
landslide studies, such as Artificial Neural Networks (ANN), neuro-fuzzy systems, and fuzzy logic methodologies. These innovative approaches facilitate the development of more nuanced and robust landslide susceptibility maps by effectively incorporating large datasets, handling uncertainty, and modelling complex relationships between variables. As a result, they can provide more accurate predictions of landslide-prone areas, helping to inform better risk management and mitigation strategies.
This study, conducted by Islik Yilmaz from the Department of Geological Engineering at Cumhuriyet University in Sivas, Turkey, was published in the Bulletin of Engineering Geology and the Environment in January 2009. It focuses on generating landslide susceptibility maps using artificial neural networks (ANN) for the Koyulhisar region, which encompasses the Boztepe, Saytepe, and Igdir Mountains, which reach elevations of 1,361 m, 1,240 m, and 1,850 m, respectively.
The study area is characterized by a geological composition that includes Pliocene volcanic deposits, the
Eocene Yesilce Formation, and Maastrichtian limestones. These older rock units are overlain by younger colluvium, which consists primarily of loose materials that have weathered from the bedrock. The oldest geological unit identified in the area comprises thinly bedded limestone that exhibits whitish-yellow and
pink coloration, with features of crushing and jointing. The Yesilce Formation is composed of a variety of materials, including conglomerates, sandy-gravelly limestone, and volcanic components such as andesites, basalts, and pyroclastics. Additionally, the Erdembaba volcanics, which are jointed and massive, include dacite, andesite, and basalt formations. The soil cover on the slopes within the study area varies in thickness, ranging from 0.5 meters to 2 to 3 meters. The average annual precipitation for this region is 394.6 millimeters, which significantly influences the area’s geological and hydrological dynamics.
The study employed a comprehensive selection of factors to assess their impact on the phenomenon
under investigation. These factors included lithology, distance from geological faults, topography (slope and aspect), elevation above sea level, distance from drainage systems, stream power index (SPI), topographic wetness index (TWI), normalized difference vegetation index (NDVI), and proximity to roads. Each of these variables was crucial for understanding the spatial dynamics involved.
To facilitate the spatial analysis, thematic layers corresponding to each of these factors were generated
using the ArcGIS software platform, which allowed for effective mapping and visualization of the spatial relationships. The analysis itself was conducted using an artificial neural network (ANN) with a back-propagation technique, which is particularly effective for modeling complex relationships in data.
For the training of the neural network, several hyper-parameters were defined to optimize the model’s
performance. Specifically, a learning rate of 0.1 was chosen, alongside a momentum parameter set at 0.9. The network was trained using a variable learning rate with momentum method, referred to as ‘traingdx,’ which helps in adjusting the learning based on the error gradient. The activation (transfer) function applied to all layers of the network was the hyperbolic tangent sigmoid function (tansig), which is commonly used for its ability to introduce non-linearity into the model.
To evaluate the neural network model’s performance, we utilized the Receiver Operating Characteristic (ROC) curve analysis along with the degree of fit (DF) metric. The results indicated that the area under the ROC curve (AUC) was 0.847, suggesting that the model has a high level of predictive accuracy. Furthermore, the degree of fit was classified as ‘Very High,’ indicating that the model’s predictions align closely with the observed data.
In addition to these evaluations, we also assessed the relative importance of the various factors included in the study. This was achieved using the frequency ratio method, which helped in determining how much each factor contributed to the modeling outcomes. The insights garnered from this analysis provide valuable information for understanding the underlying processes at play and may serve as a basis for
future studies in the field.
This study demonstrated that artificial neural networks (ANN) can effectively assess landslide susceptibility with sufficient data. However, the process is time-consuming and requires data conversion into ASCII or other formats, making it challenging to handle large datasets. The validation showed that the susceptibility map produced is of high quality, suggesting that the method can aid planners and engineers. However, it is recommended for generalized planning and assessment only.
Reference
Yilmaz, I. A case study from Koyulhisar (Sivas-Turkey) for landslide susceptibility mapping by artificial neural networks. Bull Eng Geol Environ 68, 297–306 (2009). https://doi.org/10.1007/s10064-009-0185-2