Improving daily precipitation estimation in the data scarce area by merging rain gauge and TRMM data with a transfer learning framework

• Transfer learning is introduced to improve the QPE for the data scarce area. • Fine-tuning and Domain-Adversarial Neural Network are applied in Q-T Plateau. • Domain-Adversarial Neural Network outperforms fine-tuning and co-kriging. • High similarity between target and source domains is vital for transfer learning. • Swish loss function improves performance on extreme precipitation estimation. Merging the satellite and gauge precipitation has been proofed as an efficient approach to improve the accuracy of quantitative precipitation estimation. However, for areas with few and unevenly distributed rain gauges, accurate precipitation estimation still remains challenging. To address this problem, this paper proposes a framework to improve precipitation estimation for the data scarce area based on transfer learning. Taking the Qinghai-Tibet Plateau as a representative case study, we used two transfer learning methods (fine-tuning, domain-adversarial neural network (DANN)) to transfer precipitation fusion model from the source domain to the target domain. Results indicate that in comparison with the original TRMM data, the root mean square error (RMSE) and mean absolute error (MAE) of the merged precipitation in the Qinghai-Tibet Plateau during 2001–2005 are reduced by 27.6 % and 22.5 % by using the fine-tuning method, and reduced by 29.4 % and 21.5 % by using the DANN method, respectively. Meanwhile, the correlation coefficient (CC) is increased from 0.54 (TRMM data-rain gauge data) to 0.65 (merged data-rain gauge data). The performances of the proposed methods vary spatially, with CC decreased from southeast (0.80) to northwest (<0.40) of the study area. The DANN method performed well on different precipitation intensities, while Swish loss function can help DANN achieve better results on extreme precipitation estimation, with RMSE and MAE reduced by 2.5 % and 4.5 % respectively. The performances of the proposed methods are affected by various factors such as source domain selection and the length of study period. Findings imply that transfer learning provides new insights and new methods to improve precipitation estimation for the data scarce area, which would benefit regional water-related disaster defense and water resources management.