[H-CG23] Earth surface processes related to deposition, erosion and sediment transport

[HCG23-09] Quantitative Analysis of Delta Morphologies Using Convolutional Autoencoder

*Sato Ryusei¹, Hajime Naruse¹ (1.Kyoto Univ.)

Keywords:machine learning, neural network, coastal setting, shoreline

This study quantified delta morphologies using a convolutional autoencoder and classified the morphologies. Recently, delta morphologies have been quantified and categorized to reveal the relationship with environmental factors influencing them. In previous studies, morphological features of deltas have been extracted as shoreline roughness, channel count, or presence/absence of spits. However, these metrics are scale-dependent or consider only restricted components of deltas, therefore lacking comprehensive feature detection for the wide range of delta morphologies. This study attempts to extract morphological features of deltas using a convolutional autoencoder (CAE). In this method, shoreline images of river deltas were obtained from global data of water occurrence probabilities and converted into binary images as the training dataset for CAE, which is an unsupervised learning model often used for dimension reduction or feature extraction. This study constructed a CAE model to conduct semantic segmentation for acquiring abstract features of delta morphologies. The method was applied to 80 delta images displaying various morphologies, and nodes at the middle hidden layer were extracted as abstract features of each image. The CAE model was trained effectively without serious overfitting. The morphological features of delta shorelines were summarized into 25 parameters by CAE. The k-means clustering method was applied to the extracted feature values, classifying delta morphologies into four classes. The classified delta morphologies can be interpreted as 1. bird’s foot delta, 2. arcuate delta, 3. flat shoreline delta, and 4. embayment delta. The principal component analysis implied that these four morphological classes of deltas were distributed in a morphometric space of principal components (PC) 1 and 2 without significant overlap. Class 4 had the highest PC 1 score, followed by classes 2, 1, and 3. Relative sediment fluxes were considered to examine the quantitative relationship between morphologies and environmental factors, showing a more decisive influence of river sediment transport for Classes 1 and 2. The sediment transport by waves is significant for Class 3. The correlation between tides and the morphologies could not be clearly observed. This study proposed a method using CAE and proved that it is a versatile metric for quantifying the complex shoreline morphologies of deltas. Further analysis of the adequately extracted features in various coastal settings will reveal the controlling factors for the complex landforms associated with multiple sediment transport processes, such as delta and estuary.