multilayer perceptron algorithm

Rockfalls are a type of slope movement process that does not require a transporting medium (e.g., water) and predominantly occur under the influence of gravity (Dikau, 2006). The detachment of a single rock block or a volume of blocks from steep slopes is followed by motions mainly through the air. The fall trajectory is strongly controlled by the mean slope gradient, enabling different motion modes such as free-falling, bouncing, and rolling. Consequently, rockfalls present an unpredictable threat, especially along highways and railways, where they can cause damage to infrastructure or endanger human lives (Bostjančić et al., 2020). In high mountain regions, rockfall activity is thought to be changing due to accelerated climate warming and permafrost degradation, potentially resulting in increased activity and larger volumes involved in individual falls (Stoffel et al., 2024). Therefore, it is crucial to assess areas prone to rockfalls in mountainous regions. The aim of this research is to zone the risk of rockfalls long the Khalkhal–Shahroud road using a multi-layer perceptron algorithm.

The multi-layer perceptron algorithm is a modern machine learning model capable of solving complex problems. In this study, the necessary data were obtained from topographic maps at a scale of 1:25,000, geological maps at a scale of 1:100,000, a digital elevation model (DEM) with 12.5-meter resolution from the ALOS-PALSAR satellite, Sentinel data (spatial resolution of 10 meters), Google Earth satellite images, and field studies. In rockfall hazard zoning using GIS, the most critical component is the preparation of a rockfall distribution map or rockfall inventory. Fieldwork was conducted to identify rockfalls and prepare the inventory.

To identify the factors contributing to rockfall occurrence, field studies identified eight key factors: elevation, vegetation, slope aspect, distance from faults, distance from roads, geology, land use, and slope gradient. After pre-processing, all layers were entered into SPSS Modeler software, and the model was designed with 8 input neurons, 8 intermediate neurons, and 1 output neuron. The results revealed that, in the multi-layer perceptron algorithm, the geological layer had the highest weight value (0.20), followed by the land use layer (0.14) and distance from roads (0.12). In the model validation phase, the results demonstrated an AUC value of 0.9810 in the training phase and 0.9876 in the testing phase, indicating high model validity in both phases.

Conclusion:

This research aimed to identify areas at risk of rockfall along the Khalkhal–Shahroud road using a multi-layer perceptron algorithm. The results highlight the significant influence of geological conditions on rockfall occurrences, emphasizing the need to consider slope instability in all spatial planning efforts in this region. It is recommended that future studies explore other machine learning models, such as support vector machines, to further evaluate rockfalls and related slope movements.