| Precipitation plays a significant role in the cycle of matter and energy at the globle scale. Due to the specific location, huge mountain body, and complex land surface characteristics, the Qinghai-Tibetan Plateau (QTP) has a vital impact on global climate change,as well as,is known as the 'third pole', meanwhile,it is also a critical area for monitoring and exploring the global climate change, which attracts scientists' great attentions from all of the world. The spatio-temporal information of precipitation over QTP is the important content in exploring the global climate change,and the precipitation happens over this region is one of the critical resource to recharge the five large rivers, such as Indus, Ganges, Brahmaputra, Yangtze, and Yellow river, which provide the fresh water for around 1.4 billion people in Asian.However, due to its harsh environment, remoteness, and poor transportation networks, rain gauge stations are sparse. The spatial patterns of precipitation occur with great heteroneities as influenced by the complex land surface, and the precipitation information from each station only represents a limited area (with a radius around 5 km). Therefore, how to obtain the accurate precipitation information over this region with both finer spatial and temporal scales is the fundmental and key content in the geosciences of the QTP.As the development of remote sensing technology, satellite-based precipitation products have provided much more spatiao-temporal precipitation information over the QTP, however, various challenges have come out during the development of remote sensing technology, for example, the low spatio-temporal resolutions.Therefore, one of the most key research directions is how to fuse the advantages of both satellite-basd precipitation estimates and ground observations. Currently, the optimal satellite-based precipitation product is the TMPA V7 datasets, in the TRMM era, while, the obvious shortage is the low spatial resolution (0.25°), which could not meet the various application demands for precipitation estimates with finer resolution(-1 km). This study proposed a new solution to obtain new precipitation estimates with fine resolutions (3hourly,~1 km) based on the TMPA 3B43 and 3B42 V7 datasets, and various land surface characteristics over the QTP, which concludes different downscaling algorithms at different temporal scales. The main research and conclusions are shown as follows:(1) The research on the quantative spatio-temporal relationships between precipitaton and land surface characteristics over the QTPThis study investigated the quantative spatio-tempoaral relationships between the TMPA and environmental variables over this region. The variations of precipitation and that of NDVI do not show some obvious spatial patterns at the annual scale; while NDVI responds to precipitation with a delay of 0 to 1 month at the monthly scale. The land surface temperature (LST) dominates a negative correlation with precipitation over the most areas of eastern QTP, and a positive correlation with precipitation over some areas in the northwestern QTP at annual scale; at the monthly scale, the variation of LST in the day and night could explain large variation of precipitation over most areas of QTP, except some small regions along the western edge of QTP,which shows a similar spatial patterns of the NDVI at monthly scale to simulate the precipitation, and the LST in the night shows a relatively better ability. The spatial distribution patterns of precipitation were simulated by various land surface chacteristics based on a data mining algorithm at different temporal scales. At the annual scale,~90% of the precipitation information over QTP could be explained by the land surface characteristics, while at the seasonal and monthly scales, the ability of land surface characteristics to explain the precipitation decreases. Meanwhile, we found that the land surfache characteristic with most spatial relative to precipitation varied spatially, which reveal some interesting and explainable phenomena.(2) Downscaling research based on TMPA at annual scale over the QTPTraditional downscaling strategies on TMPA data have two bottleneck problems.One is that those strategies only considered the spatial patterns of precipitation are mainly influenced by NDVI and/or NDVI; and the other is those algrithoms could not take various land surface characteristics, simultaneously, to simulate the precipitation.This study introduced a spatial data mining algorithm to solve the two bottleneck problems mentioned above, especially over the QTP with complex topography. This spatial data mining algorithm uses a divide-and-conquer strategy to downscale TMPA data over the QTP, which could better meet our assumption that the non-stationary relationships between precipitation and land surface characteristics exist with varying two-dimensional scale effects and that the relationships vary across sub-regions with different land surface variable; further, it also could select the optimal combinations of variables, which vary spatially. Another advantage of this algorithm is that it could consider various variables, simultaneously, to simulate the precipitation. All those advantages of this algorihm could meet the requirements to downscale the TMPA datasets over this region where the topography is great complex. Validating the precipitation estimates with finer resolution against the ground observations, some main conclusions were summarized as follows: (a) the accuracy of the downscaled results by Cubist was significant better, and the procedure Cubist to obtain the downscaled estimates did not need the residual corrections; (b) Cubist could divide the study area into different sub-regions, according to the geographical similarities,which considered the two-dimensional spatial scale effect; (c) in each sub-region,Cubist could select the optimal environmental variables by making the relative importantce when to simulate the co-located precipitation, which revealed some explainable information; (d) Cubist, as a data mining algorithm, has the ability to consider the thousands of variables, at the same time, to downscale the TMPA data; (e)the downscaled results by Cubist did not show any anomalies, which meant the influences of the anomalies from original TMPA data could be removed by this data mining algorithm. Therefore, this study introduced a reasonable downscaling algorithm, and it could meet the requirements for downscaling the TMPA data at annual scale over the QTP.(3) The research on removing the anomalies from the original TMPA, as well as further calibrated by the ground observationsDue to the effects of the water bodies (i.e., rivers and lakes), snow covers, there were obvious anomalies exist in the original TMPA datasets. This study also found that the simulated estimates by Cubist and the land surface characteristics demonstrated the very similar spatial patterns and accuracies with original TMPA datasets. However, the obvious anomalies from the original TMPA data could be removed from these simulated estimates. Further, the ground observations were introduced to calibrate these simulated estimates at annual scale, and then these calibrated results were disaggregated into monthly products, using geographically ratio analysis (GRA) strategy based on the original TMPA 3B43 V7 data. This new monthly datasets named by NIMTMPA3B43 QTP, from January 2000 to December 2013, has been shared on the intemet:http://agri.zju.edu.cn/NITMPA3B43 OTP/.Through validating NIMTMPA3B43 QTP against the ground observations, we found that this new monthly datasets not only remove the obviously anomalies, but also improve the accuracy, which compensate the precipitation information for monitoring the spatio-temporal precipitation change over the QTP.(4) Obtaining downscaled results of precipitation at monthly and daily scale over the QTPBased on the downscaled results at annual scale, and the assumption that the precipitation vary gradually at different spatial scales,this study proposed a new method, using spatially weighted averages based on the moving window technology to obtain the fractions on TMPA 3B43, which is used to disaggregate the annual downscaled results into monthly downscaled results. To obtain the daily downscaled results, TMPA 3B42 data (3 hourly/ 0.25°) were used to obtain the fractions, in the same way, using spatially weighted averages based on the moving window technology,and then the results were obtained with the resolution of 3 hourly/ -1 km. However, at the temporal scale of 3 hourly, the precipitation vary gradually at the temporal series,therefore, a new method, temporally weighted averages by moving window technology,was also proposed to obtain the downscaled results (3 hourly/ -1 km).Compared with the downscaling procedure at monthly scale, the downscaled results,at 3 hourly scale, considered the precipitation vary gradually both spatially and temporally. The daily downscaled results were acquired by accumulating the 3 hourly downscaled precipitation data within each day. Though validating the downscaled results at monthly and daily scales, the downscaled results not only do not have obvious anomalies but also improve the accuracy. Therefore, the new strategies proposed by this study to obtain the downscaled results at monthly and daily scales were suitable to downscale TMPA over the QTP.(5) The research on spatio-temporal change of precipitation over QTP based on the downscaled resultsFor a long time,due to the sparse ground observation network,it is difficult to obtain the spatial information of the precipitation change at the time series over the QTP. This study explored the spatio-temporal change of the precipitation based on the downscaled results at annual, seasonal, and monthly scales, over the QTP and the adjacent areas. The averages of precipitation from downscaled results and ground observations demonstrated similar trend at annual scale. And the averages of precipitation from downscaled results did not show any obvious trend at both annual and seasonal scales. However, the trend analysis based on the downscaled results demonstrated some significant spatial patterns at both annual, seasonal, and monthly temporal scales. For example, the precipitation showed decreasing trend over the southeastern QTP where the volume of precipitation is relative large influenced by India monsoon, while, precipitation showed increasing trend over the northwestern QTP where the volume of precipitation is relative small influenced by westerlies.Through validating the trend analysis from downscaled results against those from ground observations, we found that the trend analysis from downscaled results agreed well with those from ground observations with R2 around 0.83, and that at seasonal scale, only the agreement, in summer, between the trend analysis from downscaled results and ground observations is similar to that at annual scale, and the agreements decrease to around 0.70 in other seasons when the volume of precipitation relatively low. Satellite-based precipitation estimates provide much more information for monitoring the precipitation change analysis, especially at different spatial scales,which is scarcely possible by ground observations, over the QTP. Therefore, how to obtain gridded precipitation datasets with finer spatio-temporal resolutions and accuracies, as well as longer time series, is vital for investigating the precipitation change over this critical region/ the whole world.The core work of this study was that a new strategy was proposed for downscaling satellite-based precipitation estimates with low spatial resolutions,TMPA, and obtaining downscaled results(-1km) with at different temporal scales (i.e., annual, monthly, daily, and 3 hourly). Based on the research results, this study has basically arrived at the designed research goals, and the three main new progresses have been summarized in the following aspects:(1) This study introduced and improved a new spatial data mining algorithm to effectively consider the spatial relationships between the precipitation estimates,TMPA, and various land surface characteristics, and construct the models to downscale the TMPA data, which is new suitable solution for obtain downscaled results(~ 1 km) over the QTP. Meanwhile, this methodology could also be used for the worldwide scale or other regions by parameter optimization and localization.(2) Regard as the challenge on how to fuse the satellite-based and ground-based observations both spatially and temporally, this study proposed a new methodology,spatio-temporally weighted averages based on the moving window technology, to obtain the downscaled results at finer temporal scales (i.e., monthly, daily, 3huorly),by referencing the core thoughts of Geographically Weighted Regression (GWR). By validating the downscaled results at different temporal scales against the ground observations, the accuracy of the downscaled results were significantly improved, for example, the bias of monthly downscaled results was around 5%, while the bias of TMPA 3B43 V7 was around 20%.(3) Considering the original TMPA data with obvious anomalies and overestimating the precipitation,this study proposed a new methodology to remove the obvious anomalies from original satellite-based estimates and correct the satellite-based precipitation estimates with anomalies' removed using ground observations.Meanwhile, this study also provided a new monthly precipitation estimates from fusions of satellite-based and ground-based precipitation observations, from January 2000 to December 2013 (http://agri.zju.edu.cn/NITMPA3B43 QTP/). |
PDF Full Text Request