| dc.contributor | 資訊系 | |
| dc.creator (作者) | 彭彥璁 | |
| dc.creator (作者) | Chang, Chin-Ya;Peng, Yan-Tsung;Tsay, Jenq-Dar;Tsai, Tzung-Yu;Chao, Chun-Chieh;Chang, Yu-Cheng | |
| dc.date (日期) | 2026-07 | |
| dc.date.accessioned | 8-Sep-2026 15:05:12 (UTC+8) | - |
| dc.date.available | 8-Sep-2026 15:05:12 (UTC+8) | - |
| dc.date.issued (上傳時間) | 8-Sep-2026 15:05:12 (UTC+8) | - |
| dc.identifier.uri (URI) | https://ah.lib.nccu.edu.tw/item?item_id=184839 | - |
| dc.description.abstract (摘要) | Estimating rainfall from geostationary satellite observations is essential for monitoring extreme precipitation events across land and ocean surfaces, particularly in regions with complex topography. This paper describes the development and evaluation of the Artificial Intelligence Quantitative Precipitation Estimation (AIQPE) model, a deep learning framework based on a three-dimensional convolutional encoder–decoder architecture. AIQPE is designed to fuse multispectral cloud features from the Himawari-8 geostationary satellite to produce hourly precipitation estimates at 2 km spatial resolution. The main innovation of AIQPE lies in the incorporation of advanced attention mechanisms, including a Mixed Pooling Module (MPM) to capture both short- and long-range spatio-temporal dependencies, a Selective Kernel Concatenation (SK-concat) module for adaptive multimodal feature fusion, and a Ghost Bottleneck (GBN) architecture for improved computational efficiency. In addition, a custom loss function combining Mean Squared Error (MSE) and correlation is adopted to preserve rainfall structure during training. AIQPE was trained and evaluated using diverse warm-season precipitation events over Taiwan. Experimental results present substantial performance improvements over the baseline model, achieving a 22% increase in the correlation coefficient (from 0.51 to 0.63) and a 70% reduction in root-mean-square error (RMSE) (from 20.18 to 6.08). To further assess operational applicability, supplementary inter-product evaluation against IMERG and GSMaP was incorporated. The results confirm that the proposed framework provides substantial advantages for high-resolution operational rainfall monitoring under challenging meteorological and topographic conditions. To facilitate reproducibility and broader adoption, the AIQPE source code is publicly available on Zenodo at https://doi.org/10.5281/zenodo.20792000. | |
| dc.format.extent | 107 bytes | - |
| dc.format.mimetype | text/html | - |
| dc.relation (關聯) | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol.19, pp.1-11 | |
| dc.subject (關鍵詞) | Quantitative Precipitation Estimation (QPE); Geostationary Satellite; Deep Learning; 3D Convolution; Attention Mechanism; Himawari-8; Taiwan | |
| dc.title (題名) | Attention-Based Deep Fusion Convolutional Neural Network (AIQPE) for Geostationary Satellite Rainfall Estimates over Taiwan: Model Development | |
| dc.type (資料類型) | article | |
| dc.identifier.doi (DOI) | 10.1109/JSTARS.2026.3715818 | |
| dc.doi.uri (DOI) | https://doi.org/10.1109/JSTARS.2026.3715818 | |
