Quantitative Improving Precipitation Estimation For The Data Scarce Area Based On Transfer Learning

Precipitation constitutes a fundamental component of the Earth’s ecosystem,which plays a crucial role in important aspects such as water supply,agricultural production,ecosystems and climate.Investigating historical precipitation events and learning precipitation patterns within a region is of great significance for flood prevention agricultural decision-making.The actual measured values of precipitation are generally provided by hydro-meteorological stations.However,the establishment of these stations in some regions is impeded by technological,economic,and manpower constraints.Obtaining precipitation data has therefore become a significant challenge for these areas.With the global climate change,the frequency of extreme precipitation events is continuously increasing,the risk of mountain floods in data-scarce areas is becoming more severe,highlighting the urgent need for obtaining effective precipitation data.Quantitative precipitation estimation of precipitation in these regions can help address the issue of limited precipitation data and provide a basis for meteorological monitoring and water resources management.Meanwhile,improving the accuracy of precipitation estimation in data-scarce areas remains a challenging task.It has been demonstrated that fusing satellite-based precipitation data with ground-based measurements is a practical approach to improve the accuracy of quantitative precipitation estimation.However,in data-scarce areas,data fusion usually only considers one of either temporal or spatial factor,making it difficult to obtain accurate spatio-temporal features.In recent years,transfer learning has started to be applied to fields such as meteorology and hydrology.For the problem of precipitation estimation in data-scarce areas,transfer learning can be employed to extract common features between source and target domains,which expands the dataset of data-scarce areas and obtain spatio-temporal characteristics to improve the estimation accuracy.In this thesis,a transfer learning framework(fine-tuning networks and domain adversarial neural networks)was proposed to improve precipitation estimation in data-scarce regions.To demonstrate the effectiveness of the transfer learning network,this thesis focused on the Qinghai-Tibet Plateau,and combined TRMM satellite data,ground-based rain gauge data,Grid Sat-B1 data,and DEM data from 2001 to 2005.The precipitation-rich regions in China are selected as the source domain,while 82 rain gauge stations in the Qinghai-Tibet Plateau are selected as the target domain for quantitative precipitation estimation experiments.The results indicated that the proposed method can reduce the root mean square error(RMSE)and mean absolute error(MAE)by 27.6%and 22.5% with the fine-tuning network,and 29.4% and 21.5% with domain-adversarial neural networks(DANN),respectively.The use of DANN resulted in an increase in the correlation coefficient(CC)from 0.54 to 0.65.At the same time,the spatial distribution of the transfer learning results was different,with CC decreasing from the southeast(0.80)to the northwest(<0.40)of the study area.The DANN method performed well under various precipitation intensities,and the Swish loss function can help DANN achieve better results in extreme precipitation estimation,with RMSE and MAE reduced by 2.5% and 4.5%,respectively.Given that the performance of transfer learning is affected by factors such as source domain selection and length of the study period,this thesis further discussed the 1999-2019 fusion results’ performance using data from 51 rain gauges,and generated a 21-year fusion dataset for the Qinghai-Tibet Plateau region.The thesis suggested that transfer learning provides novel insights and methods for enhancing precipitation estimation accuracy in data-scarce regions,which can benefit regional water resource management,disaster prevention,and agricultural production.