Transfer Learning with a Convolutional Neural Network for Hydrological Streamline Detection

Keywords: Transfer Learning, Convolutional neural network, Remote Sensing, Streamline detection
Session Type: Virtual Paper Abstract
Day: Saturday
Session Start / End Time: 2/26/2022 02:00 PM (Eastern Time (US & Canada)) - 2/26/2022 03:20 PM (Eastern Time (US & Canada))
Room: Virtual 3

Abstract

Streamline network delineation plays a vital role in various scientific disciplines and business applications, such as agriculture suitability, river dynamics, wetland inventory, watershed analysis, surface water survey and management, flood mapping. Traditionally, flow accumulation techniques have been used to extract streamline, which delineates streamline primarily based on topological information. Recently, machine learning techniques such as the U-net model have been applied for streamline detection l. In this paper, we examine the usage of transfer learning techniques, transfer the knowledge from a prior area and use the knowledge of the prior area as the starting point of the model training for a target area. We also tested transfer learning methods with different scenarios, changed input data by adding the NAIP dataset, retrained the lower and the higher part of the network, and varying sample sizes. We use the original U-net model in the previous research as the baseline model and compare the model performance with the model trained from scratch. The results demonstrate that even though the transfer learning model leads to better performance and less computation power, it has limitations that need to be considered.