Patterns of Geomorphic Processes Across Deltas Using Image Analysis and Machine Learning

The morphology of deltas is determined by the spatial patterns of geomorphic processes that continuously modify the landscape through erosion and deposition. While generally resilient, deltas are threatened by changing environmental conditions such as sea level rise and increased storm intensity that can drastically alter the rates and patterns of sediment transport. We present a method for extracting information about the nature and spatial extent of active geomorphic processes in deltas from the geometry of islands and the channels around them. By identifying island clusters with similar morphology and link these to dominant physical processes, we can predict how a delta system might respond to changes in its driving forcings.

Using the Ganges-Brahmaputra-Meghna (GBM) Delta as a case study, we applied image processing methods to delineate islands on a water mask generated from satellite imagery and measured various geometric properties of islands and channels. These morphological metrics were converted into linearly uncorrelated parameters using principal component analysis (PCA). We then used three unsupervised machine learning algorithms in sequence to cluster islands in this dataset into spatially continuous zones of similar morphology: (1) GeoSOM, an artificial neural network that considers both the geometric attributes and spatial proximity of polygons; (2) a Gaussian Mixture Model (GMM), and (3) a radius-based Nearest Neighbor Classifier (NNC) to aggregate results.

We find that the GBM Delta can be divided, without user input, into 6 major zones that correspond to expected distributions of geomorphic processes across the delta. This classification shows good correspondence with the known extent of the tidally-influenced region of the delta and indicates that the impact of fluvial processes is limited to narrow corridors along the Ganges and Hoogly rivers. It does not recognize human-made 'polders' or the natural mangrove forest of the Sundarbans as separate morphological classes from surrounding islands, suggesting that they are geometrically similar to their neighbors. These results demonstrate that machine learning and remote-sensing imagery are useful tools for automatically identifying the spatial patterns of geomorphic processes across delta systems.