Journal of Hydrology: Regional Studies (Feb 2024)

Identifying private pumping wells in a land subsidence area in Taiwan using deep learning technology and street view images

  • Chun-Wei Huang,
  • Si Ying Yau,
  • Chiao-Ling Kuo,
  • Tsai-Yu Kuan,
  • Si-Yu Lin,
  • Ching-Shih Tsou,
  • Chuen-Fa Ni,
  • Yuan-Chien Lin,
  • Liang-Cheng Chang

Journal volume & issue
Vol. 51
p. 101636

Abstract

Read online

Study region: The Choushui River Fan, Taiwan. Study focus: Groundwater overdraft has led to not only groundwater depletion but also environmental disasters, such as subsidence and seawater intrusion in the Choushui River Alluvial Fan, Taiwan. The influence of land subsidence is gradually shifting from the coast to the center of the fan and threatening Taiwan high-speed rail. However, it remains a great challenge to manage and model the groundwater aquifer due to numerous unregulated wells. This study maps and locates private wells using deep learning technologies. We trained and validated convolutional-based deep learning neural networks (DNNs), using street view images. We applied the DNNs to a land subsidence area along the Taiwan high-speed rail, termed the Golden Corridor in Taiwan. The results showed that DNNs can recognize pumping wells with at least 90% accuracy. The testing cases showed their capability to recall all the pumping wells in three road segments along the Golden Corridor. Finally, we spatially estimated potential pumping of a subsidence area using the fine-trained DNNs. New hydrological insights for the region: Given the prevalence of unknown private pumping in the Choushui River Fan, our image data-driven computer vision approach not only eases labor-intensive private well investigations but also advances hydrologic understanding for groundwater modeling. We enhance comprehension of unknown sinks and provide their spatial distribution to improve groundwater modeling.

Keywords