matlab code of TSEB_DNN

Evapotranspiration (ET) and its components of soil evaporation (E) and vegetation transpiration (T), as key variables for the water-energy exchange between the land surface and the atmosphere, are widely used in hydrological and agricultural applications. The land surface temperature based two-source energy balance (TSEB) model can provide high accuracy E, T, and ET, which are spatio-temporally discontinuous, whereas the spatio-temporally continuous daily ET is more helpful in water resources management. In this study, to improve the continuity of estimates from the TSEB model, we developed a new combined model coupling the TSEB model and deep neural network (DNN) (TSEB_DNN). First, spatio-temporally continuous reference data was prepared based on the remote sensing and meteorological data as input, and E from soil and T from vegetation were obtained from the TSEB model under clear-sky condition as outputs. Then, the DNN was trained under clear-sky condition to obtain the relationship between E and T estimates from TSEB and reference data. Finally, the trained DNN was driven by the spatio-temporally continuous reference data to obtain spatio-temporally continuous E, T, and total ET. Compared with the ET estimates from the original TSEB model, the continuity was significantly improved for the TSEB_DNN model. The TSEB_DNN model was well consistent with the in situ measurements and had overall correlation coefficient (R), root-mean-square-error (RMSE), and bias values of 0.88, 0.88 mm d-1, and 0.37 mm d-1, respectively. The ratio of T/ET estimates from the TSEB_DNN model had high accuracy against in situ measurements with RMSE and bias values of 7.49% and -2.22%, respectively. The combined model and the maps of E, T, and ET will help improve water resource management.

蒸散发（Evapotranspiration, ET）及其组成部分土壤蒸发（soil evaporation, E）与植被蒸腾（vegetation transpiration, T）是实现地表与大气之间水热交换的关键变量，被广泛应用于水文与农业研究领域。基于地表温度的双源能量平衡（two-source energy balance, TSEB）模型能够提供高精度的E、T与ET估算结果，但此类估算结果存在时空不连续性问题；而时空连续的日尺度蒸散发数据对水资源管理更具实用价值。本研究为改善TSEB模型估算结果的连续性，提出了一种耦合TSEB模型与深度神经网络（deep neural network, DNN）的新型组合模型（命名为TSEB_DNN）。具体研究流程如下：首先，以遥感与气象数据为基础构建时空连续的参考数据集，并以晴空条件下TSEB模型输出的土壤蒸发E与植被蒸腾T作为训练标签；随后，在晴空条件下训练DNN，以学习TSEB模型输出的E、T与参考数据之间的映射关系；最后，利用时空连续的参考数据集驱动训练完成的DNN，从而得到时空连续的E、T及总蒸散发ET。相较于原始TSEB模型的ET估算结果，TSEB_DNN模型的估算结果连续性得到显著提升。该模型与原位实测数据吻合度良好，整体相关系数（correlation coefficient, R）、均方根误差（root-mean-square-error, RMSE）及偏差（bias）分别为0.88、0.88 mm·d⁻¹与0.37 mm·d⁻¹。TSEB_DNN模型输出的T/ET比值估算结果同样具有较高精度，其RMSE与bias分别为7.49%与-2.22%。本研究提出的组合模型及对应的E、T、ET空间分布图将有助于提升水资源管理水平。