River Morphology Classification Using Deep Convolutional Neural Networks

Abstract

River morphology classification has been researched since the 19th century. With the rise of higher-resolution satellite imagery and the development of more advanced machine learning algorithms, automated and procedural classification can be employed. This thesis investigates to what extent a (Deep) Convolutional Neural Network ((D)CNN) can be applied on satellite imagery to perform river morphology classification and identifies the optimal architecture with a limited dataset. This research represents the first attempt at using deep learning methods to classify river morphologies and demonstrates the feasibility and potential of DCNNs. Several experiments are employed to find the most suitable model for river classification. The optimal architecture is a two-input DCNN utilizing satellite images and centerlines of the rivers of the SWOT River Database as input, having five convolutional and pooling layers, with 24, 48, 96, 192, and 384 feature maps in the convolutional layers. It includes a fully connected layer with 64 dense units, dropouts of 0.2 after each pooling layer and between the fully connected- and the output layer, and batch normalization. The weighted F1-score is 53% but the network still exhibits slight overfitting despite the use of regularization techniques. The final model was able to classify the four different river morphologies, anastomosing, braiding, meandering, and wandering, with an F1-score of 25%, 49%, 37%, and 65%, respectively. The DCNN significantly outperforms a baseline that predicts the majority class. A Cohen’s d value of 0.9 demonstrates the DCNN's large improvement over the baseline. The model has difficulty distinguishing between anastomosing and braided and meandering and wandering which is to be expected since human classifiers also perceive this difficulty. Additionally, most river segments have multiple labels in one image, which makes the classification challenging. This study was the first to explore and demonstrate the feasibility and potential of using DCNNs to classify river morphologies. Gathering more labeled data is necessary for further improvement of the network's performance.