Scholarlink

Near real-time satellite precipitation data with high accuracy is crucial for water resource management, short-term weather forecasting, and disaster preparation. Currently, there are many near real-time satellite precipitation products available; however, their accuracy is still limited compared to delayed-time precipitation products. This paper developed a hybrid deep learning model, referred to as DeepModel, combining U-Net and ResNet. It uses AWS data, ERA-5 atmospheric reanalysis data, and ASTER digital elevation model data to calibrate the IMERG-Early Run precipitation product for Vietnam. The proposed model achieved classification metrics (POD: 0.69, FAR: 0.5) and regression metrics (CC: 0.435, CV: 1.716) better than the delayed-time products IMERG-Final Run and GSMaP-MVK, as well as the near real-time products IMERG-Early Run and PERSIANN-CCS. The paper’s results open up prospects for employing various deep learning techniques and additional auxiliary data, such as weather radar, microwave sensors, and numerical weather prediction, to enhance the accuracy of near real-time precipitation products.

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

Calibration of Near Real-Time Satellite Precipitation Products Using Deep Learning Techniques and Multi-source Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Anh-Duc Hoang-Gia,
  • An Hung Nguyen,
  • Truong Xuan Ngo,
  • Thanh Thi Nhat Nguyen

摘要

Near real-time satellite precipitation data with high accuracy is crucial for water resource management, short-term weather forecasting, and disaster preparation. Currently, there are many near real-time satellite precipitation products available; however, their accuracy is still limited compared to delayed-time precipitation products. This paper developed a hybrid deep learning model, referred to as DeepModel, combining U-Net and ResNet. It uses AWS data, ERA-5 atmospheric reanalysis data, and ASTER digital elevation model data to calibrate the IMERG-Early Run precipitation product for Vietnam. The proposed model achieved classification metrics (POD: 0.69, FAR: 0.5) and regression metrics (CC: 0.435, CV: 1.716) better than the delayed-time products IMERG-Final Run and GSMaP-MVK, as well as the near real-time products IMERG-Early Run and PERSIANN-CCS. The paper’s results open up prospects for employing various deep learning techniques and additional auxiliary data, such as weather radar, microwave sensors, and numerical weather prediction, to enhance the accuracy of near real-time precipitation products.