Development Of Multi-source Remote Sensing Based Evapotranspiration Models For Natural Ecosystems And Urban Areas

Soil evaporation(E)and vegetation transpiration(T)are important components of the land surface water cycle and energy cycle.Accurate estimation of E and T at a regional scale have important theoretical and practical significance for understanding the impact of climate change on the hydrological cycle,regional drought prediction,agricultural precision irrigation,and guidance for regional water resources allocation and management.Traditional measurements and simulations of E and T are carried out on a single point scale.In recent years,with the development of remote sensing technology(RS),remote sensing images combined with land surface physical and physiological processes based models have been used to simulate E and T in regional scale.This study systematically compared two types of dual-source models based on thermal infrared remote sensing component temperature decomposition:(1)an iterative process based component temperatures decomposition model(Two-Source Energy Balance Model,TSEB);(2)a land surface temperature-fractional vegetation coverage trapezoid framework based(T_m-f_c)component temperatures decomposition model(Modified Pixel Component Arranging and Comparing Algorithm,M-PCACA);as well as an impedance process based dual-source model(Penman-Monteith-Mu,PM-MU)in estimation of the E and T in China.The three models were verified on six cloudless test days in 2015 with 21 flux observation sites including 7 underlying surfaces and one isotope station in the Heihe River Basin.The verification results show that the M-PCACA model performs best(RMSE=38.7 W/m~2),followed by the TSEB model(RMSE=44.5 W/m~2),and the worst performance is the PM-MU model(RMSE=61.4W/m~2).The results of further verifying the three models on the 7 underlying surfaces show that the TSEB model has certain advantages in areas with high vegetation coverage and high leaf area index(LAI)(such as forest areas).The uncertainty of the TSEB model mainly comes from:(1)The determination of the initial P-T value and the use of the one P-T value in different underlying surface areas;(2)The iterative target of soil evaporation greater than zero leads to a systematic overestimation of ET and T;(3)The empirical P-T coefficient lacks a clear physical mechanism.Due to the characteristics of the T_m-f_c trapezoid,the M-PCACA model has most stable model performance in various underlying surface types.Therefore,it is recommended to use in the surface conditions with a high degree of heterogeneity,but the uncertainty of the M-PCACA model mainly comes from:(1)Determination of the extreme dry and wet boundary of the T_m-f_c trapezoid;(2)Underestimation of the latent heat flux(LE)in most cases caused by the neglect of advection.The PM-MU model performs better in some non-vegetation coverage areas(such as deserts,water bodies,etc.),but performs poorly in vegetation coverage areas.The uncertainty of the PM-MU model mainly comes from:(1)There is large error in the use of relative humidity and water vapor pressure to characterize soil moisture changes;(2)The determination of the empirical parameterβin complementary related theories leads to systematic errors.Although the overall performance of the PM-MU model is not as good as the two types of models based on the thermal infrared temperatures,the PM-MU model does not require the input of surface temperature,therefore the PM-MU model is very suitable for continuous regional E and T simulation and produce of ET data sets.In order to further evaluate the impact of dry and wet boundary location in T_m-f_ctrapezoid framework to regional evapotranspiration simulation,three types of T_m-f_ctrapezoid extreme dry and wet boundary location methods were compared:(1)The method based on the actual boundary of the T_m-f_c trapezoid(AET);(2)The method based on the theoretical boundary of T_m-f_c trapezoid(TET);(3)The method of coupling the actual boundary and theoretical boundary of T_m-f_c trapezoid(ATET).The results of the three methods show that changing the four extreme temperatures will significantly change the simulation of LE,but will not change the simulation of the ratio of vegetation transpiration to evapotranspiration(LE_v/LE).Comparing the verification results of 11 flux stations in southwestern China shows that different extreme temperature combinations change the root mean square error(RMSE),Bias and mean absolute error(MAPD)of the LE simulation but do not change the coefficient of determination(R~2).In addition,the different extreme temperature combinations will not change the spatial patterns of LE and LE_v/LE.The results of comparing the three methods show that ATET performs better than AET in various underlying surface types,and when the underlying surface heterogeneity is higher,the performance gap between the two methods is further shrink and vice versa.The TET method exhibits unstable performance in six test days period.TET is better than the ATET on some test days,but worse than the AET on some test days.However,a significant advantage of TET is that it can be applied in various size of the underlying surface especially can be used in very large-scale.Except for the very large underlying surface type,it is recommended to use the ATET method to locate the theoretical dry and wet boundary of the T_m-f_c trapezoid.In order to verify the performance of the proposed PCALEP model for simulating E and T,the improved Shuttleworth-Wallace(S-W)model and two types of machine learning models(random forest and support vector regression)were used to partition the ET components in a cabbage farmland in the suburbs of Yueyang City,Hunan Province.The improved S-W model takes into account the impact of carbon dioxide concentration and emissions on vegetation transpiration simulation,and proposes the Monte Carlo method to estimate the empirical parameters in the calculation of soil surface impedance,which significantly improves the accuracy of the original S-W model.Results suggest compared with original S-W model,the R~2 can be increased from 0.73 to 0.80,and RMSE can be reduced from 0.54 mm/d to 0.41 mm/d.In addition,it was the first attempt to use a machine learning model to predict vegetation transpiration and achieved good simulation results.For random forest and support vector regression,there is the best simulation accuracy when all meteorological data,vegetation data and soil data are input.But when the air temperature and solar radiation are lacking,both two types of machine learning models show the obviously underestimated and poor accuracy for vegetation transpiration simulation.Comparing the two models results show that when the input data is relatively complete or only lack of soil and vegetation data,it is recommended to use the random forest model;if the basic meteorological data(temperature and radiation)are missing,the support vector regression model was performs better.Based on the M-PCACA model by Zhang et al.(2005),in order to overcome the problems in the M-PCACA model,the PCALEP model is further proposed.The PCALEP model assumes that vegetation and nearby soil have different energy drives and different water responses.Soil evaporation is controlled by the water content of the surface soil,while vegetation transpiration is controlled by the root soil water content.In addition,PCALEP uses a newly T_m-f_c trapezoid actual extreme boundary automatic positioning method TBLA(trapezoidal boundary location algorithm).TBLA develops different positioning algorithms according to the different scattered point distributions of the dry and wet boundaries of the T_m-f_c trapezoid,the method fully considering the relatively sparse,scattered and uneven scatter distributions in wet boundary of T_m-f_ctrapezoid and significantly improves the reliability of the actual boundary location.The verification results with four flux sites in southwestern China show that the PCALEP model is more reliable for LE simulation accuracy.The RMSE and Bias simulated at Yueyang site are 36.3 W/m~2 and-21.3 W/m~2,respectively;RMSE and Bias are 39.7W/m~2 and-27.7 W/m~2,respectively at Xishuangbanna site.In addition,the comparison LE_v/LE results with S-W model results in Yueyang site and two types of machine learning models proves that PCALEP can provide accurate estimates of E and T.The PCALEP model simulates the LE_v/LE of the study area in 2015 is 0.69,which proves that T is dominant in total ET.Although the dual-source remote sensing model for ecosystems is flourish,the multi-source remote sensing model for urban areas is still scarce.The above-mentioned two source remote sensing models perform large error for simulating ET in urban areas.By coupling the model idea based on the thermal infrared remote sensing component temperature decomposition and the idea based on the canopy stomatal impedance,proposed the first urban three-source remote sensing evapotranspiration model(TRU).TRU first divides the urban underlying surface into impervious areas and non-impervious areas based on patch model theory and urban land use information.In the simulation of ET in non-impervious areas,the parameterization scheme uses soil surface temperature to characterize soil moisture content and estimates soil evaporation based on complementary relationship theory,and develops a new iterative method for decomposing component temperature.The evaporation of the impervious area is estimated independent using the humidity gradient.The model was tested for 45cloudless days in Tianjin for 2017-2020 based on 30 m Operational Land Imager(OLI)/Enhanced Thematic Mapper Plus(ETM~+)images.Verification results with the two EC towers in urban suburban woodland area and main urban area show that the model can provide a more reliable LE simulation.The RMSE was 38.8 W/m~2,the MAPD was 11.4%,and Bias was 9.9 W/m~2 for LE simulation on 45 test days.For the LE_v/LE simulation,the RMSE verified with the stable isotope is 0.087,the MAPD is12.3%,and the Bias is-0.012.Comparison with urban single-source models and two-source models for ecosystem suggest TRU provide best accuracy for ET and its components simulation.The spatial pattern suggested impervious surface evaporation exhibited minimal seasonal variation and maintained a very lower level due to limited availability of water.The results emphasized the importance of using land use and land cover(LULC)in urban ET modeling and the necessity to calculate ET as independent of impervious areas.TRU represents a groundbreaking development of multi-source urban satellite-based ET models and facilitates the mapping of urban ET components.The new algorithm proposed in this study has significantly improved the accuracy of RS based ET simulation in ecosystems and urban areas,and in-depth analysis the differences of model performance and physical mechanisms of different two source models on various underlying surfaces.This study provides a good theoretical basis and reference for the selection of model in regional ET simulation,and points out the direction for future model improvement.Based on the research in this article,the future improvement of RS-based multi-source models for ecosystems and urban areas could focus on the following aspects:(1)Coupling ideas based on thermal infrared temperatures and canopy impedance,the fusion of RS thermal infrared,near infrared and visible light to develop a spatial and temporal continuous regional E and T data sets;(2)Improve the accuracy of key RS products for estimating regional E and T,such as thermal infrared surface temperature,soil moisture,vegetation height products,etc.;(3)To enhance the using of machine learning models in regional RS-based ET models to simplify model operations and generate E and T datasets;(4)Multi-source ET models in urban areas need to be further developed.The currently researchers most focus on the development of ET models in ecosystems,but urban areas as gathering places for human activities require more attention;(5)Expend the surface flux observation of the fluxnet sites,especially the ground sites in urban areas are still very scarce,and there is an urgent need to establish ground ET observation data sets in urban areas.