Transfer Learning with Convolutional Neural Networks for Hydrological Streamline Delineation

Keywords: Deep learning, Transfer Learning, Convolutional neural network, Remote sensing, Streamline detection, Terrain analysis
Abstract Type: Paper Abstract

Abstract

Streamline network delineation is essential for various applications, such as agriculture sustainability, river dynamics, and watershed analysis. Recently, machine learning methods have been applied for streamline delineation and have shown promising performance. However, their performance drops significantly when a trained model is applied to different locations. In this paper, we explore whether fine-tuning neural networks that have been pre-trained on large label datasets (e.g., ImageNet) can improve transferability from one geographic area to another. Specifically, we collected smaller label datasets of the National Hydrography Dataset (NHD) from the Rowan County, NC and Covington River, VA areas in the eastern United States. First, we fine-tune pre-trained ImageNet models on the Rowan County area and compare them with an attention U-net model that is trained from scratch on the same dataset. We find that the DenseNet169 model achieves an F1-score of 85% which is about 4% higher than the attention U-net model. Then, to compare the transferability of the models to a new geographic area, we select the three highest F1-score models from the Rowan County area and further fine-tune them with the data in the Covington area. Similarly, we fine-tune the attention U-net model from the Rowan County area with the data in the Covington area. We find that fine-tuning ImageNet models achieve an F1-score of 71.87% in predicting the stream pixels in the Covington area, which is significantly higher than training the models from scratch in the Covington area or fine-tuning attention U-net model from Rowan to Covington.