ارزیابی تطبیقی الگوریتم‌های یادگیری ماشین و یادگیری عمیق در استخراج لندفرم‌های مناطق بیابانی (مطالعه موردی: جنوب خاوری گرمسار)

Extended Abstract

Introduction

Geomorphological mapping plays a central role in understanding landform patterns, surface processes, and landscape evolution, and it underpins a wide range of applications including natural hazard assessment, desetif-

ication studies, land-use planning, and envi-

ronmental management. Although tradition-

al manual interpretation has long been the dominant approach, it is time-consuming, subjective, and highly dependent on expert experience. Advances in satellite remote sensing, high-resolution elevation data, and automated analysis methods have substanti-

ally expanded the potential for producing geomorphological maps with greater effeci-

ency and reproducibility.

Machine learning techniques particularly

Random Forest have been widely adopted for geomorphological classification and have demonstrated strong performance in distinguishing landforms with clear spectral or morphometric signatures. However, their pixel-based nature limits effectiveness in areas with complex textures, heterogeneous patterns, or overlapping spectral characteris-

tics. Deep learning models, especially U-Net architectures, offer a promising alternative due to their ability to learn spatial context, texture, and object-level structures. Recent studies have shown their superiority in mapping landforms such as dunes, alluvial fans, and agricultural terraces.

Despite these advancements, limited resear-

ch has integrated multispectral Sentinel-2 imagery with morphometric and spectral indices within a unified deep learning framework for desert geomorphology. This study addresses this gap by evaluating the performance of U-Net, in comparison with Random Forest, for producing detailed geomorphological maps in an arid environ-

ment with diverse landform units.



Methodology

This study employed Sentinel-2A imagery from 2024 (13 spectral bands at 10–60 m resolution), combined with morphometric indices including total curvature, the Topographic Wetness Index (TWI), terrain roughness (Gourabi, 2023), and the Normalized Difference Moisture Index (NDMI). Two classification approaches were applied to generate the geomorphological map: the Random Forest algorithm and the U-Net deep learning architecture.

Random Forest, a supervised and non-parametric ensemble model, constructs multiple bootstrap-sampled decision trees using both dependent variables (geomorphological classes) and independent variables such as slope and curvature. The algorithm reduces tree-to-tree correlation through random feature selection and aggregates predictions through majority voting.

U-Net, originally designed for biomedical image segmentation, is a fully convolutional neural network with symmetric downsampling and upsampling paths. Its encoder–decoder structure enables pixel-level classification and the precise extraction of complex landform boundaries, making it highly suitable for heterogeneous desert landscapes.

To build the training dataset, field surveys were conducted and GPS-based samples were cross-checked with existing geomorphological information. After integrating all datasets, both models were trained and implemented in Python.

Finally, an independent set of field validation samples was collected to evaluate model performance using Overall Accuracy and the Kappa coefficient.





Results and discussion

The landform classification produced by the Random Forest algorithm identified five geomorphic units and seventeen landforms using a combination of Sentinel-2 reflective bands, the normalized moisture index, and morphometric parameters. Among these classes, dissected hillocks and wetland areas represent the largest spatial extent, covering approximately 372 km² and 382 km², respectively. In contrast, the playa-related alluvial fan exhibits the smallest area, with about 5 km². The results also show that hogbacks occupy around 185 km² in the southern part of the study area, while active and inactive gravelly plains cover roughly 66 and 58 km² in the northeastern and northern sectors.

Using the U-Net deep learning model, five units and sixteen landforms were mapped; however, this approach was unable to differentiate new alluvial fans from older ones. The largest class corresponds to dissected hillocks with 349 km². Alluvial fans, floodplains, and clay flats cover approximately 13%, 12%, and 11% of the study area. Gypsum outcrops form the smallest landform, with only 0.4 km². Active and inactive gravel plains are also scattered across the region, primarily overlying alluvial fan surfaces.

For accuracy assessment, 500 field-verified control points were used. The Random Forest model achieved an overall accuracy of 89% and a Kappa of 0.86, while U-Net showed superior performance with 94% overall accuracy and a Kappa of 0.92. Conclusion

Producing geomorphological maps is a complex process influenced by data type and classification method. In this study, Sentinel-2A imagery was combined with morphometric indices—including the Topographic Wetness Index, total curvature, terrain roughness—and the Normalized Difference Moisture Index to compare the performance of Random Forest and U-Net in mapping desert landforms. The comparison showed that each method possesses distinct strengths and limitations. Random Forest performed better in distinguishing landforms with clear boundaries and strong morphometric signatures, such as young and old alluvial fans. This advantage is likely related to its sensitivity to variables such as slope and curvature. U-Net, however, achieved higher accuracy in landforms characterized by complex spatial textures, heterogeneous patterns, and nonlinear boundaries, including transitions between agricultural land, salt-affected areas, fluvial deposits, and gravel plains. Its ability to learn spatial texture, neighborhood structure, and contextual patterns allowed U-Net outputs to more closely reflect real-world conditions. A key finding is that U-Net mapped patchy and fragmented landforms more accurately, whereas the pixel-based nature of Random Forest tended to homogenize them. Random Forest also showed reduced accuracy where spectral overlap was high such as clay plains and wet areas while U-Net delineated boundaries more effectively. Moreover, Random Forest struggled with linear anthropogenic features like roads, requiring manual correction, whereas U-Net extracted them more reliably. Overall, Random Forest is more suitable for spectrally and morphometrically distinct landforms, while U-Net performs better in heterogeneous units. A hybrid approach could therefore produce more realistic geomorphological maps in future studies.