بازسازی بلندمدت الگوهای بادی از طریق تحلیل لندفرم‌های بادی ریگ یلان شرق دشت لوت، با استفاده از یادگیری ماشین

Abstract

Introduction

Tucked away in the eastern reaches of Iran’s Lut Desert, Rig-e Yalan sprawls across roughly 9,800 square kilometers, standing as one of the world’s most arid and vibrant desert landscapes. With annual rainfall barely trickling past 50 millimeters and vegetation so sparse it’s almost a mirage, this region is sculpted by relentless regional winds that etch out a mesmerizing array of aeolian landforms—towering dunes, sweeping wind corridors, and restless sand ridges that seem to shift with every gust. These dunes are like nature’s archives, their shapes, orientations, and intricate patterns telling stories of the winds that have shaped them over countless years. My goal in this study was to dive into these stories, mapping the dominant wind directions and unraveling the geomorphological dance of Rig-e Yalan using a blend of cutting-edge machine learning tools—Random Forest and K-Means clustering. What excites me most about this approach is how it bypasses the need for exhaustive field campaigns, offering a fresh, scalable way to decode desert dynamics that could inspire similar explorations across the globe.

Rig-e Yalan’s elongated, oval form, stretching about 150 by 70 kilometers along a northeast–southwest axis, feels like a living canvas, painted by a symphony of regional weather systems—monsoonal winds, local airflows—and the rugged terrain that channels them. The sun blazes down here, with summer temperatures often soaring past 45°C, and the air is so parched it seems to pull the moisture right out of you. These extreme conditions fuel a relentless erosional force, giving rise to a stunning variety of dunes: linear ones that stretch like ribbons across the horizon, crescent-shaped barchans that glide over the sands, and complex star dunes that defy simple description. I drew inspiration from studies in places like China’s Hobq and Kumtag Deserts or Africa’s Namib Desert, where researchers have shown how dune shapes can unlock secrets about past climates and wind patterns. But what’s less common—and what I set out to tackle—is using a fusion of machine learning and clustering to map wind patterns across an entire region. This work bridges that gap, offering a deep dive into how winds and landforms intertwine, with insights that could resonate far beyond Iran’s borders.

To pull this off, I turned to advanced remote sensing, pulling data from the Shuttle Radar Topography Mission (SRTM) and high-resolution satellite imagery to extract 15 geomorphometric indices—think of them as the landscape’s vital signs, revealing its structure and behavior. These allowed me to pinpoint wind directions by analyzing dune shapes without ever stepping foot in the desert, a method that feels almost like reading the land from afar. I grouped these directions into eight main categories: north, northeast, east, southeast, south, southwest, west, and northwest. The results don’t just deepen our understanding of Rig-e Yalan’s wind-driven world; they open doors to practical applications—tackling climate change impacts, managing wind erosion, picking prime spots for wind farms, planning military operations, or safeguarding fragile desert ecosystems in places where data is scarce. By leaning on technology rather than costly field treks, this approach is both practical and far-reaching, potentially guiding sustainable resource management in deserts worldwide, from the Sahara to the Gobi.

Materials and Methods

This study unfolded through a carefully crafted seven-step journey, designed to peel back the layers of Rig-e Yalan’s aeolian landscape: pinpointing the study area, gathering data, cleaning it up, extracting geomorphometric features, analyzing dune shapes, modeling wind patterns with machine learning, and clustering similar regions. I chose Rig-e Yalan for its extreme aridity and diverse landforms—its dunes, from crescent-shaped barchans to sprawling linear ones, are like a natural laboratory for studying wind patterns. The region’s hyper-arid climate, with its sparse vegetation and relentless winds, makes it a perfect canvas for exploring how landscapes evolve under the influence of aeolian forces.

The data came from a rich mix of sources: Landsat 8/9 and Sentinel-2 satellite imagery, offering 10–30-meter resolution, SRTM and ALOS PALSAR digital elevation models at 30-meter resolution, and Google Earth for close-up visual checks of dune shapes. Preprocessing was like tuning an instrument before a concert—I used techniques like Dark Object Subtraction (DOS), FLAASH, and 6S to correct for atmospheric noise, ensuring the data was as clear as the desert sky. From the elevation models, I extracted 15 geomorphometric indices: slope (SLP), aspect (ASP), general curvature (CURV), plan curvature (PC), profile curvature (PRC), height variation (HV), topographic position index (TPI), surface roughness (RUG), sediment transport index (STI), terrain ruggedness index (TRI), slope direction gradient (SDIR), windward/leeward index (WI), local elevation difference (ED), local relief (LR), and relative position index (RPI). These metrics were my tools for decoding how dunes move and how winds shape them, each one offering a piece of the puzzle.

For modeling wind patterns, I chose Random Forest with 50 decision trees—a method I admire for its reliability and ability to wrestle with complex, non-linear data without tripping over itself. I trained it using wind direction data from 3,948 field stations, sorting directions into those eight main classes. To group similar landscape zones, I turned to K-Means clustering, settling on six clusters after testing with the Elbow and Silhouette Score methods, which helped me find the sweet spot for grouping without forcing unnatural divisions. This clustering approach felt like sorting the desert into distinct neighborhoods, each with its own topographic character.

Validation was crucial to ensure the results held up under scrutiny. I used statistical measures like overall accuracy (around 78%), Kappa coefficient (about 0.64), F1-Score, mean absolute error (MAE), and mean squared error (MSE). For a reality check, I compared the model’s predictions to actual dune shapes seen in Google Earth imagery, looking for that moment of alignment between data and desert. The toolkit was a powerhouse: Google Earth Engine for crunching remote sensing data, ArcGIS and SAGA for spatial analysis, Python (with scikit-learn, TensorFlow, and XGBoost) for the machine learning heavy lifting, and R and MATLAB for stats and visualizations. Together, these tools wove a seamless pipeline for dissecting Rig-e Yalan’s geomorphological story, blending precision with creativity.

Discussion and Results

The findings painted a vivid picture of Rig-e Yalan’s wind-driven world: southeast (SE) and south (S) winds dominate, shaping 41.42% and 39.59% of the landscape, respectively, while east (E) and northeast (NE) winds play a quieter role, influencing just 19% of the area. The Random Forest model nailed the wind direction classifications, hitting an accuracy of about 78% and a Kappa of 0.64. It shone brightest for SE, S, and SW directions, with F1-Scores above 0.8, showing it’s got a sharp eye for the main players. The west (W) direction was a tougher nut to crack, with a lower F1-Score of 0.31, likely because its dune shapes blend into neighboring classes or get tangled in local terrain quirks. When I cross-checked the model’s predictions with real-world dune patterns in Google Earth, I found a solid 78% match, especially in areas with crisp barchan and linear dunes in the south and center—moments when the data felt like it was singing in harmony with the landscape.

K-Means clustering revealed six distinct morphogenic zones, each with its own topographic personality. Zones 1 and 3, with gentle slopes (5–10 degrees) and smoother surfaces, were hotbeds of dune movement, swept along by the dominant winds. Zones 2 and 4, tucked in leeward areas with rougher terrain (roughness > 0.5), were more stable, acting as sediment traps where the sands settle. I also spotted anomalies—519 pixels, or about 129.75 square kilometers, that didn’t follow the dominant wind patterns, mostly clustered in the central and eastern parts. These outliers likely stem from secondary eastern winds or localized terrain quirks, like sudden ridges or dips that nudge the winds off course.

The geomorphometric indices were the heart of this analysis. The average slope (11.54 degrees) suggested a mostly gentle landscape, but steeper slopes (up to 77.4 degrees) in certain spots steered the winds like natural funnels. The aspect, averaging 174.6 degrees, showed slopes mostly facing south-southeast, perfectly aligned with the main wind directions. The sediment transport index (mean 3.97) highlighted intense sediment movement in steep, rough areas—places where the desert is in constant motion. Southern regions with negative curvature were erosion hotspots, where winds scour the land, while central and northern areas with positive curvature were where sediments piled up, forming natural repositories.

The Shannon entropy map lit up the central and southern regions as the most complex (entropy > 1.5), signaling high morphodynamic activity and erosion risk—areas where the desert is alive with change. These findings echo studies in deserts like Hobq and Kumtag in China or the Namib in Africa, proving that machine learning can unlock wind patterns without needing boots on the ground. The blend of Random Forest and K-Means clustering let me pinpoint homogeneous zones and spot anomalies with precision, like finding hidden patterns in a vast, sandy tapestry. Statistically, elevation had little to do with dune movement (R² = 0.0299), confirming winds as the main sculptor. Indices like topographic position and sediment transport correlated positively with dune shifts (r > 0.6), while roughness tied negatively to stability in sheltered areas (r < -0.4), painting a clear picture of how the landscape responds to wind.

Conclusion

By weaving together Random Forest and K-Means clustering, this study offers a fresh lens on Rig-e Yalan’s aeolian landforms and long-term wind patterns. The Random Forest model, with its 78.36% accuracy and 0.64 Kappa, proved a trusty guide for predicting dominant winds, while K-Means clustering mapped out six zones that mirrored the region’s wind and topographic rhythms. This approach, needing minimal field data, delivered a vivid picture of aeolian dynamics, backed up by Google Earth comparisons that felt like a nod from the desert itself.

The study isn’t without its challenges—long-term field data is hard to come by, the eight-direction classification might miss finer nuances, and remote sensing resolution has its limits. I’d love to see future work place portable sensors in anomaly-prone areas, like the central and eastern zones, and blend field data with numerical models to sharpen the picture. Extending this method to other Iranian dune fields, like Rig-e Jen or the Central Kavir, could unravel broader regional wind patterns and climatic shifts, adding new chapters to the story.

The implications are vast: from grappling with climate change and wind erosion to picking ideal wind farm sites, planning military strategies, or safeguarding desert ecosystems. This framework is a blueprint for data-driven geomorphological studies in arid regions, offering a scalable, cost-effective way to inform sustainable resource management and paleoclimate research globally. It’s a method that feels alive with possibility, ready to explore deserts from the Sahara to the Gobi and beyond.

Keywords: Rig-e Yalan, LUT Desert, Geomorphometry, Machine Learning, Wind Patterns, Random Forest, K-Means