Scholarlink

<p>Assessing groundwater sustainability is challenging where monitoring data are limited or discontinuous, making it harder to draw reliable conclusions. Groundwater levels in shallow, unconfined aquifers are influenced by surface hydrology, anthropogenic land use practices, and recharge processes (e.g., indirect recharge from rivers and lakes). This study introduces a convolutional neural network (CNN) model that estimates groundwater levels from surface water spatiotemporal dynamics represented by the normalized difference water index (NDWI) derived from Sentinel-2 satellite imagery. The CNN model was trained using historical groundwater levels observed in two wells in unconfined aquifers in the Midwestern USA from 2018 to 2023 and corresponding NDWI images of the surrounding areas. A total of 305 and 291 NDWI images (filtered for &lt; 25% cloud coverage; area 5120 m<sup>2</sup>; resolution 10&#xa0;m × 10&#xa0;m) were used for the two sites, respectively. The CNN model achieved low root mean square error (RMSE) values of 0.25&#xa0;m and 0.18&#xa0;m and high Pearson and Spearman correlations (≥ 0.90) for the two sites. Furthermore, three typical machine learning (ML) methods were tested: K-nearest neighbors, random forests, and support vector machines, which provided reasonable results when combined with the principal component analysis (minimum RMSE values of 0.30 and 0.31&#xa0;m for the two sites, respectively). However, the CNNs outperformed the ML models. Regardless of the ML algorithm employed, this study illustrates the ability to predict groundwater levels using satellite-derived NDWI.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Modeling groundwater levels using satellite images and machine learning

  • 期刊论文

  • 发表日期：

    2025年10月13日

查看全文
  • Nikolaos Nagkoulis,
  • Mohammad Shamsudduha,
  • Emmanouil Anagnostou,
  • Amvrossios Bagtzoglou

摘要

Assessing groundwater sustainability is challenging where monitoring data are limited or discontinuous, making it harder to draw reliable conclusions. Groundwater levels in shallow, unconfined aquifers are influenced by surface hydrology, anthropogenic land use practices, and recharge processes (e.g., indirect recharge from rivers and lakes). This study introduces a convolutional neural network (CNN) model that estimates groundwater levels from surface water spatiotemporal dynamics represented by the normalized difference water index (NDWI) derived from Sentinel-2 satellite imagery. The CNN model was trained using historical groundwater levels observed in two wells in unconfined aquifers in the Midwestern USA from 2018 to 2023 and corresponding NDWI images of the surrounding areas. A total of 305 and 291 NDWI images (filtered for < 25% cloud coverage; area 5120 m2; resolution 10 m × 10 m) were used for the two sites, respectively. The CNN model achieved low root mean square error (RMSE) values of 0.25 m and 0.18 m and high Pearson and Spearman correlations (≥ 0.90) for the two sites. Furthermore, three typical machine learning (ML) methods were tested: K-nearest neighbors, random forests, and support vector machines, which provided reasonable results when combined with the principal component analysis (minimum RMSE values of 0.30 and 0.31 m for the two sites, respectively). However, the CNNs outperformed the ML models. Regardless of the ML algorithm employed, this study illustrates the ability to predict groundwater levels using satellite-derived NDWI.