TY - JOUR
T1 - Advancing Satellite-Derived Precipitation Downscaling in Data-Sparse Area Through Deep Transfer Learning
AU - Zhu, Honglin
AU - Zhou, Qiming
N1 - This work was supported in part by the Natural Science Foundation of China (NSFC) General Program under Grant 41971386 and Grant 42271416 and in part by the Hong Kong Research Grant Council (RGC) General Research Fund under Grant HKBU 12301820.
PY - 2024/2/19
Y1 - 2024/2/19
N2 - Improving the spatial resolution of satellite-based precipitation data is crucial in expanding the application of quantitative precipitation estimation. Many machine-learning (ML)-based techniques have been proposed for the downscaling of satellite-derived precipitation. However, these methods often highly rely on the availability of observation data, posing limitations on their applicability in data-scarce regions. To address this issue, a novel approach based on deep transfer learning (TL) was introduced to improve the precipitation downscaling in data-scarce areas. The proposed framework implemented the fine-tuning to transfer the pretrained convolutional neural network (CNN)-based downscaling model from the source domain to two target domains. Results showed that compared with the original satellite data, the correlation coefficient (CC) of the downscaled precipitation using fine-tuning increased to 0.715 and 0.622 in the two target domains, respectively. The transferred models with three fine-tuned layers achieved the highest performance, and models with one fine-tuned layer were the most efficient. The simplest transferred models with four frozen layers obtained the worst performance (with the CC values of 0.586 and 0.519 in two target domains). By incorporating at least one fine-tunable layer, the TL models demonstrated significant improvements. Meanwhile, the transferred downscaling models had better performance under higher precipitation intensity and worse for low precipitation intensity. They received more accurate downscaled results during the wet season than the dry season. This study provided a promising transfer leaning-based approach in generating high-resolution precipitation in data-sparse regions, which has great significance to the meteorological, hydrological, and ecological research in such areas.
AB - Improving the spatial resolution of satellite-based precipitation data is crucial in expanding the application of quantitative precipitation estimation. Many machine-learning (ML)-based techniques have been proposed for the downscaling of satellite-derived precipitation. However, these methods often highly rely on the availability of observation data, posing limitations on their applicability in data-scarce regions. To address this issue, a novel approach based on deep transfer learning (TL) was introduced to improve the precipitation downscaling in data-scarce areas. The proposed framework implemented the fine-tuning to transfer the pretrained convolutional neural network (CNN)-based downscaling model from the source domain to two target domains. Results showed that compared with the original satellite data, the correlation coefficient (CC) of the downscaled precipitation using fine-tuning increased to 0.715 and 0.622 in the two target domains, respectively. The transferred models with three fine-tuned layers achieved the highest performance, and models with one fine-tuned layer were the most efficient. The simplest transferred models with four frozen layers obtained the worst performance (with the CC values of 0.586 and 0.519 in two target domains). By incorporating at least one fine-tunable layer, the TL models demonstrated significant improvements. Meanwhile, the transferred downscaling models had better performance under higher precipitation intensity and worse for low precipitation intensity. They received more accurate downscaled results during the wet season than the dry season. This study provided a promising transfer leaning-based approach in generating high-resolution precipitation in data-sparse regions, which has great significance to the meteorological, hydrological, and ecological research in such areas.
KW - Convolutional neural network (CNN)
KW - downscaling
KW - fine-tuning
KW - precipitation
KW - transfer learning (TL)
UR - http://www.scopus.com/inward/record.url?scp=85186110131&partnerID=8YFLogxK
U2 - 10.1109/TGRS.2024.3367332
DO - 10.1109/TGRS.2024.3367332
M3 - Journal article
AN - SCOPUS:85186110131
SN - 0196-2892
VL - 62
JO - IEEE Transactions on Geoscience and Remote Sensing
JF - IEEE Transactions on Geoscience and Remote Sensing
M1 - 4102513
ER -