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

ABSTRACT
Geomorphological maps are essential tools in analyzing geomorphological, hydrological processes, and natural resource management. Traditional methods for producing these maps, which rely on field observations and visual interpretation of aerial images, face limitations due to being time-consuming and costly. two approaches—machine learning (Random Forest) and deep learning (U-Net)—were used to prepare an accurate geomorphological map of southeastern Garmsar. The input data included Sentinel-2A images from 2024, a 10-meter Digital Elevation Model, and morphometric indices such as Topographic Wetness Index, general curvature, roughness, and normalized moisture index. The results exhibited that the Random Forest algorithm had significant performance in distinguishing landforms with pronounced morphometric features, particularly young and old alluvial fans. In contrast, U-Net provided higher accuracy in identifying landforms with complex textures and irregular boundaries, such as hills, clay plains, wetlands, saline lands, and active and inactive sandy plains. The U-Net algorithm also reconstructed the boundaries between mountainous and plain units with greater delineation. The overall accuracy and Kappa coefficient of Random Forest were 89% and 0.86, respectively; whereas U-Net indicated an overall accuracy of 94% and a Kappa coefficient of 0.92. Overall, the results reveal that the integrated application of machine learning and deep learning algorithms, combined with remote sensing data and morphometric indices, has high capability in precise extraction of landforms, especially in arid environments with spatial heterogeneity and complex boundaries. This approach can serve as an efficient and generalizable framework for geomorphological mapping, environmental change monitoring, and supporting decision-making in sustainable land management.
Extended Abstract
Introduction
Geomorphological maps, providing spatially classified landform and surface process information, are fundamental tools for environmental analysis and spatial planning (Bishop et al., 2012; Paron & Claessens, 2011). Traditional map production is time-consuming, costly, and heavily dependent on expert interpretation, limiting reproducibility and generalization (Lark et al., 2014; Randle et al., 2018). Advances in high-resolution remote sensing, digital elevation models, and computational tools have enabled semi-automated and automated approaches, allowing detailed mapping of both natural and anthropogenic landforms (Evans, 2012; Giaccone et al., 2022). Machine learning algorithms, particularly Random Forest, have been widely applied for landform classification due to their ability to integrate diverse topographic and environmental variables, even with limited training data (Rodriguez-Galiano et al., 2012; Veronesi & Hurni, 2014). Meanwhile, deep learning, - especially convolutional networks like U-Net - allows direct extraction of spatial and textural patterns from imagery (Zhao et al., 2025; Li et al., 2025). This study aims to map arid landforms in southeastern Garmsar by combining Sentinel-2A imagery with morphometric and spectral indices and comparing the performance of Random Forest and U-Net in detecting complex boundaries and heterogeneous landforms in desert environments.
 
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 utilizing Overall Accuracy and the Kappa coefficient.
 
Results and Discussion
Landform classification was conducted using Sentinel-2A reflectance data, normalized moisture index, and morphometric indices from the digital elevation model. Both U-Net and Random Forest successfully identified five major units: mountains, hills, alluvial plains, playas, and anthropogenic areas. U-Net recognized 16 primary landforms, while Random Forest identified 17, including slopes, erosion-affected hills, gypsum and salt outcrops, hogbacks, old and young alluvial fans, active and inactive sand flats, riverine deposits, clay plains, wetlands, saline lands, and agricultural areas.
Random Forest delineated old (126 km², 36%) and new (54.8 km², 2.7%) alluvial fans more accurately, whereas U-Net combined them into a single class (274.4 km², 13.7%). Conversely, U-Net provided finer boundaries for erosion-affected hills and alluvial fans (349.6 km², 17.5%) and better spatial continuity between agricultural and saline lands, likely due to its deep network structure. Random Forest misclassified some areas as wetlands. Both algorithms detected active and inactive sand flats, but Random Forest produced contiguous patches, while U-Net captured scattered patterns that matched field observations.
Floodplains, clay plains, and riverine deposits were better delineated by U-Net, particularly for the Ab Dolati River course. The differences reflect algorithmic approaches: Random Forest is pixel-based, while U-Net integrates spatial, textural, and neighborhood patterns. Validation with 100 ground control points showed higher overall accuracy for U-Net (90%, Kappa 0.87) than Random Forest (88%, Kappa 0.85), with class-specific accuracies highlighting variations in landform recognition.
The study’s outputs, in order to assess the capability of modern methods in landform mapping, were compared with the classical 1:500,000 geomorphological map of Iran, which provides only general units due to its small scale. Using 10 m resolution Sentinel-2A imagery combined with morphometric indices (curvature, terrain roughness, topographic wetness), both Random Forest and U-Net enabled more detailed landform delineation. Random Forest performed well for landforms with distinct morphometric differences, while U-Net better captured complex boundaries and spatial heterogeneity. The results highlight that integrating remote sensing with morphometric indices and machine learning or deep learning allows higher-resolution, more precise geomorphological mapping than traditional methods.
 
Conclusion
Producing geomorphological maps is a complex process that depends on the type of input data and the classification methods employed. The present study aimed to evaluate the combined application of machine learning and deep learning approaches for landform identification using Sentinel-2A imagery, morphometric indices (including topographic wetness index, general curvature, and terrain roughness), and normalized moisture in arid environments. Since the input data were identical for both algorithms, differences in the resulting maps can be mainly attributed to the nature of the algorithms and their classification mechanisms. The results should not be interpreted as evidence of the absolute or generalizable superiority of one algorithm over the other. Algorithm performance depends on landform type, spectral overlap, boundary complexity, and the spatial scale of patterns. Although overall accuracy and Kappa indices provide a general assessment, class-level accuracy metrics demonstrate that the comparative advantage of each method differs among geomorphological units, precluding the assumption of consistent performance across classes. Random Forest exhibited weaker performance in detecting linear human-made features such as roads, requiring manual correction and pixel adjustments, which may introduce human bias. In contrast, U-Net extracted these features with higher accuracy without extensive intervention. Despite resampling all datasets to a 10 m resolution, intrinsic mismatches between spectral and morphometric data remain, as each pixel represents different information, and small landforms or complex boundaries may be smoothed or homogenized. This limitation reduces the generalizability of the results to other regions and scales. Overall, the findings indicate that the targeted application of machine learning and deep learning algorithms - tailored to landform type and study scale - can serve as an effective tool for producing accurate geomorphological maps for land-use planning, water resource management, erosion control, and environmental hazard assessment in arid areas. Developing hybrid, multi-scale approaches leveraging the strengths of each algorithm represents a promising direction for forthcoming studies.
 
Funding
There is no funding support.
 
Authors’ Contribution
Authors contributed equally to the conceptualization and writing of the article. All of the authors approved thecontent of the manuscript and agreed on all aspects of the work declaration of competing interest none.
 
Conflict of Interest
Authors declared no conflict of interest.
 
Acknowledgments
We are grateful to all the scientific consultants of this paper.