Document Type : Research Paper

Authors

• Shadi Jalilian ¹
• Shaban Shataee Jouibary ²
• Mohammad Hadi Moayeri ²
• Amir Saddodin ³

¹ Ph.D. Student, Department of Forestry, Faculty of Forest Sciences, Gorgan University of Agricultural Sciences and Natural Resources, Gorgan, Iran

² Professor, Department of Forestry, Faculty of Forest Sciences, Gorgan University of Agricultural Sciences and Natural Resources, Gorgan, Iran

³ Professor, Department of Watershed Management, Faculty of Rangeland and Watershed Management, Gorgan University of Agricultural Sciences and Natural Resources, Gorgan, Iran

Abstract

Introduction
Landslides, as one of the most destructive natural disasters, cause significant annual human and financial losses worldwide, particularly in Golestan Province due to its specific topographic and climatic conditions. Although numerous studies have been conducted in this region, most have relied on classical and traditional statistical methods, leaving a notable gap in the application of advanced deep learning algorithms. This research aimed to address this gap with three main innovations: 1) employing deep learning models (CNN and RNN) for the first time in the region, 2) conducting a systematic comparison with classical machine learning models (RF and SVM), and 3) focusing on a forest watershed with a sensitive and complex ecosystem (Hyrcanian). 
 
Materials and methods
The study area was Forest Watershed 85, covering approximately 39,288 hectares in Golestan Province. Based on previous studies and analysis of regional conditions, ten factors influencing landslide occurrence were selected: elevation, slope percentage, slope aspect, Topographic Wetness Index (TWI), Normalized Difference Vegetation Index (NDVI), land use, distance to roads, distance to rivers, distance to faults, and average annual rainfall. These factors were prepared as raster layers with a 30x30 meter cell size in a Geographic Information System (ArcGIS) environment. Landslide point data (247 points) were obtained from the General Department of Natural Resources and Watershed Management of Golestan Province. To balance the dataset, an equal number of points (247) were randomly selected from areas without landslide phenomena. Thus, a balanced dataset comprising 494 points was prepared for modeling. The data were split into a 70:30 ratio (346 points for training and 148 points for validation). Four advanced models, including two machine learning models (Random Forest, RF, and Support Vector Machine, SVM) and two deep learning models (Convolutional Neural Network, CNN, and Recurrent Neural Network, RNN), were implemented and trained in the Python environment using the Scikit-learn and TensorFlow libraries. The performance of the models was quantitatively evaluated using the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC) index. Furthermore, the validation of the generated susceptibility maps in high-risk areas (high and very high-risk classes) was performed using 30% of the unused samples, assessed by the overall accuracy and Kappa coefficient metrics.
 
Results and discussion
In this research, the performance of four machine learning and deep learning algorithms was evaluated for landslide susceptibility zoning in Forest Watershed 85 of Golestan Province. The results demonstrated that the deep learning algorithm CNN, with an AUC score of 0.910, an overall accuracy of 87.72%, and a Kappa coefficient of 0.899 for high-risk classes (high and very high risk), was identified as the most efficient model. Variable importance analysis using the superior model (CNN) revealed that the factors Distance to Fault and Distance to River were respectively the most significant contributors to landslide occurrence in the study area. This finding is entirely consistent with the geological and geomorphological expectations of the region. The final landslide susceptibility map generated by the CNN model indicated that 31.46% of the watershed area (approximately equivalent to 12,360 hectares) is classified within the very high-risk category.
 
Conclusion
The findings of this research clearly confirm the superiority and high potential of deep learning algorithms, particularly the Convolutional Neural Network (CNN) architecture, for producing high-accuracy landslide susceptibility maps compared to machine learning algorithms. The generated map can serve as a reliable and powerful scientific tool for managers and planners, enabling them to prioritize preventive measures, risk management, and safe land-use planning by focusing on high-risk areas. For future studies, integrating these models with optimization algorithms, utilizing higher-resolution data to overcome the limitations of the current data, and developing hybrid frameworks that enhance both accuracy and interpretability are recommended.

Keywords

• Convolutional Neural Network (CNN)
• Hyrcanian forests
• Landslide susceptibility prediction
• Receiver Operating Characteristic (ROC) curve
• Spatial modeling