Spatially Transferable Hydrologic Streamline Delineation: A Meta-Learning Approach

Keywords: hydrologic streamline delineation, meta-learning, Model-Agnostic Meta-Learning (MAML), geographic information systems (GIS), convolutional neural networks (CNNs).
Abstract Type: Paper Abstract

Abstract

Hydrologic streamline delineation is critical for effective water resource management and environmental monitoring. Traditional methods, such as flow accumulation from Digital Elevation Models (DEMs), face limitations due to spatial heterogeneity, complex connectivity, and high computational demands. Recent convolutional neural network (CNN) approaches, like the attention U-Net, have improved accuracy but require extensive retraining to generalize across diverse regions. This study introduces a meta-learning approach using Model-Agnostic Meta-Learning (MAML) to enhance model adaptability and scalability across varied geographic contexts. We compared four training strategies—joint training, meta-learning, training from scratch, and fine-tuning—using preliminary tests in Covington, where the meta-learning model achieved performance comparable to other strategies while requiring only 25 samples per episode, significantly reducing data needs and training time from the usual 1400 samples. Further testing across 623 watersheds in Alaska confirmed these findings, with meta-learning demonstrating similar precision, recall, and F1-scores to conventional methods. This study highlights meta-learning’s effectiveness in addressing spatial variability with minimal retraining, establishing it as a promising, resource-efficient approach for large-scale hydrological applications across diverse landscapes.