Identification and classification of desert landforms using the random forest algorithm in the Garmsar region

Introduction

Geomorphological maps provide detailed insights into landforms, surface processes, and terrain evolution, and have been widely developed across the world. These maps are not only of scientific importance but also serve essential roles in natural hazard assessment, urban planning, archaeological surveys, land use management, and climate change adaptation. Traditional methods for geomorphological mapping—based on fieldwork and manual interpretation of topographic maps and aerial photos—are often time-consuming, costly, and subjective. Over the past three decades, advancements in remote sensing and digital elevation models have enabled the development of semi-automated and quantitative mapping techniques. Among these, machine learning algorithms such as Random Forest have shown high performance in supervised landform classification. This study aims to produce a detailed geomorphological map of the arid regions of Dehnamak and Aradan using Sentinel-2A data and the Random Forest algorithm. The region has not been the subject of previous similar studies, making this research a valuable contribution to high-precision landform mapping and the broader application of advanced classification techniques in arid environments of Iran.





Methodology

This study aims to classify landforms in a mountainous and arid region using the Random Forest (RF) algorithm and to assess the impact of integrating morphometric indices with satellite imagery on classification accuracy. The study area is located on the southern slopes of the Central Alborz Mountains, overlooking the Central Iranian Plateau. Geographically, it spans parts of Semnan and Tehran provinces, including mountainous terrains in the north and desert areas in the south, mainly situated between Semnan and Garmsar counties. Sentinel-2A imagery was used as the primary remote sensing dataset. Additionally, three key morphometric indices—Topographic Wetness Index (TWI), Curvature, and Surface Roughness—were derived from a Digital Elevation Model (DEM) to improve terrain characterization. Landform classification was conducted in two stages: first, using only Sentinel-2A imagery with the RF algorithm; and second, by combining the morphometric indices with the Sentinel-2A data in the RF model. Accuracy assessment was performed using the Kappa coefficient and Overall Accuracy metrics.



Results and Discussion

The analysis of landform classification results using two distinct approaches—a spectral model based solely on Sentinel-2A data and a combined model integrating morphometric parameters (curvature and surface roughness)—revealed significant differences in the accuracy and quality of landform identification. Statistical and spatial outputs from both models showed varying patterns of coverage and separability across geomorphological classes.

Certain classes such as agricultural lands, mountainous areas with shallow valleys, eroded mountain slopes, and fluvial deposits exhibited similar classification accuracies in both models. For instance, the area of agricultural lands was estimated at 122.5 km² (4.6%) in the spectral model and 109.2 km² (4.1%) in the combined model, indicating minimal difference due to their distinct spectral features and relatively simple topography.

Conversely, classes like young alluvial fans, clay plains, and salt flats showed better accuracy in the spectral model. For example, young alluvial fans covered 397.5 km² (15%) in the spectral model but only 318.5 km² (12%) in the combined model. The salt flats also showed a sharp drop in the combined model—from 99.1 km² (3.7%) to 27.5 km² (1%)—due to reduced sensitivity to spectral brightness caused by the emphasis on morphometric features.



In contrast, the combined model performed better in identifying complex geomorphic units such as hills, regular mountain slopes, and hogbacks. Quantitative validation using 100 random ground control points showed higher accuracy for the combined model (85% overall accuracy, Kappa = 0.82) compared to the spectral model (78%, Kappa = 0.74). These findings confirm that integrating spectral and morphometric data improves landform classification in topographically complex environments and aligns with prior studies (e.g., Regmi et al., 2024; Veronesi & Hurni, 2014).



Conclusion

Landform mapping is a complex process influenced by data type and classification methods. This study evaluated the performance of the Random Forest algorithm using two scenarios: one based solely on Sentinel-2 spectral data (optical model), and another combining spectral data with morphometric indices—Topographic Wetness Index (TWI), curvature, and roughness (combined model). Results showed that integrating spectral and morphometric data improved classification accuracy for certain landforms, although not uniformly across all classes.

While both models performed similarly for units such as agricultural land, shallow-slope mountains, playa margins, and badlands, the optical model yielded better results for classes like salt flats, clay plains, and new alluvial fans—highlighting the strength of spectral data in distinguishing units with unique reflectance. Conversely, the combined model outperformed in identifying landforms like undulating hills, floodplains, hogbacks, and structured mountains, where topographic variation is more significant.

Overall, the combined model increased overall accuracy from 78% to 85% and the Kappa index from 0.74 to 0.82, demonstrating improved landform delineation. This suggests that combining spectral and morphometric variables provides a more robust classification, especially in geomorphologically diverse areas. Future improvements may involve using multi-temporal data, deep learning methods, and optimized variable integration.