Unravelling the evolution of alluvial fans in the northern Sultanate of Oman: applications of remote sensing and Deep Learning

Alluvial fans are distinctive fluvial landforms that develop at the abrupt widening of mountain valleys due to decreasing slope. They can occur in various environments and geomorphological settings (Ventra & Clarke, 2018). Their formation is influenced by the interplay between tectonics and climatic conditions. Alluvial fans are conical landforms formed along a drainage system where a topographic gradient exists and deposition prevails over erosion. In such context, a fan results from aggrading sediments that spread out from a sedimentary source through multiple channels radiating from the apex, which may shift over time (Blair & McPherson, 2009). The activation/deactivation of channels ultimately depends on the local hydrological regime. As a consequence, alluvial fans preserve evidence of various generations of inactive channels (paleochannels), lateral migrations of active streams, and the establishment of relative chronology based on their geometrical relationships. In the northern Sultanate of Oman, the southern margin of the Al-Hajar Mountains is flanked by extensive alluvial fans, forming vast and coalescing bajada-type landforms. The semi-arid to arid climate of the region, combined with the almost complete absence of plant cover, allows the observation of the bare surface of alluvial fans, which exhibits intricate and complex patterns of (paleo-)drainage systems. The latter consist of a series of exhumed gravel ridges, representing alluvial fan systems with ages ranging from the Miocene to the Pleistocene (Blechschmidt et al., 2009). Notably, previous works (Maizels, 1987, 1990; Maizels & McBean, 1990) has identified up to 14 generations of paleochannels along the alluvial fan at Barzaman. The current availability of multispectral high-resolution SPOT 6 and 7 satellite imagery offers unprecedented opportunity to investigate these paleochannel systems across an area of approximately 1780 km2. Understanding these landforms is critical for reconstructing past hydrological conditions of the Barzaman alluvial fan, which occurrence at the foothills of mountain belt reflects the role of climatic and tectonic influence on its evolution. However, manual mapping is time-consuming, and the subjective interpretation of fluvial features and paleochannels can lead to inconsistencies, characterised by different level of generalisation, in the assessment of hydrological history and landscape evolution. To address these challenges, we propose implementing Machine Learning/Deep Learning techniques to automate the detection of fluvial features and the tracing of paleochannel paths, thereby enhancing mapping accuracy and consistency, ultimately facilitating a better understanding of alluvial fan dynamics. Preliminary results involve the manual mapping of paleochannels and fluvial features in a portion of the area using 4-bands SPOT satellite imagery, with a resolution of 6 m. This approach enables the recognition of several generations of paleochannels based on geometric relationships, including intersections and overlaps. The mapped area serves as dataset for developing algorithms that leverage Deep Learning techniques, specifically employing Convolutional Neural Networks (CNNs). This step allows for the production of a probabilistic map for identifying paleochannel systems, differentiating between paleochannels and the underlying substrate. CNN models are particularly effective for the segmentation of alluvial fans from multispectral imagery due to its capacity to capture both spatial and spectral features, which are essential for accurate identification and mapping. CNNs are capable of learning the spatial context, given the sequence of convolution operators, that distinguishes alluvial fans from other landforms. Furthermore, they have the capacity to generalize effectively across diverse geographic regions and environmental conditions. These characteristics enable the model to automate the segmentation process over extensive datasets. In particular, a U-Net architecture (Ronneberger et al., 2015) was selected for the purpose of alluvial fan segmentation. The U-Net model's encoder-decoder architecture is particularly well-suited for image segmentation tasks, where the accurate delineation of boundaries is of paramount importance. The U-Net model is capable of capturing both low-level and high-level features due to the presence of contracting (encoder) and expansive (decoder) paths, which progressively reduce and then recover spatial resolution. A comprehensive learning pipeline (PyTorch) was developed based on the U-Net architecture, which entails the mapping of fluvial landforms in analogous environmental contexts. In the specific, the preliminary dataset comprising 387 multispectral images, of which 270 images were augmented and allocated for training over 50 epochs, and 117 images were used for testing. The model's performance was evaluated using accuracy metrics, achieving 93% accuracy on the testing set. This research is expected to yield a high-resolution geomorphological map of the Barzamani alluvial fan and develop Deep Learning algorithms for the automated tracing and identification of fluvial and geomorphological features. The development of this approach will enhance the potentiality for remote mapping of alluvial fans and other fluvial landforms in semi-arid and arid environments, in order to significantly advance the understanding of alluvial fan dynamics and provide valuable insights for future research in fluvial geomorphology. We kindly acknowledge the support of ESA for providing the access to the SPOT imagery through project PP0100418 (A. Pezzotta), as well as to the Erasmus + Traineeship 2024/25 grant for enabling A. Pezzotta the opportunity to conduct this research at Charles University.