陆表潜热通量是地表土壤、植被水分蒸散发所导致的从地表传递到大气的能量。潜热通量是连接全球水和能量平衡的关键参量。高空间分辨率的地表潜热通量估算对区域农业水管理、区域环境监测、局部气候变化监测具有重要的意义。但由于高空间分辨率遥感数据重访周期长、易受云污染等特点所造成的数据缺失，并且现有高分辨率遥感潜热通量产品都是使用单一的算法，这使得现有高分辨率潜热通量产品具有较大的不确定性与误差。

在人工智能技术日渐成熟的背景下，利用深度学习方法融合多种高分辨率潜热通量产品，提高高分辨率潜热通量产品精度成为重要的研究问题。本文使用convolutional neural
network-long short-term memory（CNN-LSTM）融合算法整合了五种高分辨率遥感潜热通量产品、地形信息（高程、坡度、坡向）和EC地面观测数据，建立深度学习融合框架，进而生成了一种高空间分辨率、高精度的遥感潜热通量产品。同时，本文利用该产品对多个尺度的潜热通量进行空间制图，分析其时空变化特征。本文主要的研究结论如下：

（1）本研究使用地面观测数据对五种使用单一算法的高分辨率遥感潜热通量产品（RS-PM、SW、MS-PT、PT-JPL和UMD-SEMI产品）进行验证。验证数据为来自FLUXNET的实测数据，其站点数据分布在全球各地，共190个站点，时间跨度为2001-2015年。验证结果表明，没有一种遥感潜热通量产品在所有的地物类型上表现都是最好的。这说明使用单一算法的产品适用的地表条件不同，在大尺度区域并没有最优的算法。

（2）在对五种产品验证的基础上，本研究使用多种机器学习与深度学习的方法对产品进行融合，进而得到最优的遥感潜热通量产品。其中，CNN-LSTM融合算法结合了两种深度学习模型（CNN与LSTM），可以充分的利用输入数据的空间与时间信息。除了CNN-LSTM融合算法，本研究还选用了其他三种传统的机器学习融合算法，包括多重线性回归方法（multiple
linear regression, MLR）、随机森林（random forest, RF）和深度神经网络（deep neural networks, DNN）。交叉验证结果表明，CNN-LSTM融合算法可以有效的提高潜热通量估算的精度，相比于单一的遥感潜热通量产品和其他机器学习融合方法，RMSE降低了2-8 W/m²并且KGE提高了0.04-0.16。CNN-LSTM融合算法考虑了周围像元的影响，并生成了16天平均的遥感潜热通量产品，它是一种基于先验知识的算法。这种方法可以提供更可靠的高分辨率遥感潜热通量产品，进而观测地表异质性较高区域的水文变量。

        The latent heat flux (LE) is the energy transferred from the surface to the atmosphere caused by the evapotranspiration of the surface soil and vegetation. LE is a key parameter linking global water and energy balances. The estimation of LE with high spatial resolution is of great significance for regional agricultural water management, regional environmental monitoring, and local climate change monitoring. However, due to the long revisit period of high spatial resolution remote sensing data and the lack of data caused by cloud contamination, and the existing high-resolution remote sensing LE products all use a single algorithm, which makes the existing high-resolution remote sensing LE products have large uncertainties and errors.

In the context of the increasingly mature artificial intelligence technology, the use of deep learning methods to integrate multiple high-resolution LE products and improve the accuracy of high-resolution latent heat flux products has become an important research issue. This paper uses the convolutional neural network-long short-term memory (CNN-LSTM) fusion algorithm to integrate five high-resolution remote sensing LE products, terrain information (elevation, slope, aspect) and EC ground observation data to establish deep learning. The fusion framework is then used to generate a high-spatial-resolution, high-precision remote sensing latent heat flux product. At the same time, this paper uses this product to map the latent heat flux at multiple scales and analyze its temporal and spatial variation characteristics. The main research contents of this paper mainly include the following three points:

(1) This study uses ground observation data to validate five high-resolution remote sensing LE products (RS-PM, SW, MS-PT, PT-JPL, and UMD-SEMI products) using a single algorithm. The verification data is the measured data from FLUXNET, and its site data is distributed all over the world, with a total of 190 sites, and the time span is 2001-2015. Validation results show that no single remote sensing LE product performs best on all feature types. This shows that products using a single algorithm are applicable to different surface conditions, and there is no optimal algorithm in large-scale areas.

(2) Based on the verification of five products, this study uses a variety of machine learning and deep learning methods to integrate the products, and then obtain the optimal remote sensing LE product. Among them, the CNN-LSTM fusion algorithm combines two deep learning models (CNN and LSTM), which can fully utilize the spatial and temporal information of the input data. In addition to the CNN-LSTM fusion algorithm, this study also selected three other traditional machine learning fusion algorithms, including multiple linear regression (MLR), random forest (RF) and deep neural network (DNN). The cross-validation results show that the CNN-LSTM fusion algorithm can effectively improve the accuracy of LE estimation. Compared with a single remote sensing LE product and other machine learning fusion methods, the RMSE is reduced by 2-8 W/m2 and the KGE is improved 0.04-0.16. The CNN-LSTM fusion algorithm considers the influence of surrounding pixels and generates a 16-day averaged remote sensing latent heat flux product, which is an algorithm based on prior knowledge. This method can provide a more reliable high-resolution remote sensing latent heat flux product, which in turn enables observation of hydrological variables in areas with high surface heterogeneity.

  (3) This study uses the CNN-LSTM fusion remote sensing LE product to map at different scales, including farmland scale, regional scale, and global scale. At the farmland scale, the spatial distribution of the remote sensing LE products fused by CNN-LSTM is similar to other single remote sensing LE products, but there are differences in the average size. On the regional scale and the global scale, the spatial distribution of LE of CNN-LSTM fusion is reasonable, and the average value is also within a reasonable range, which indicates that the fusion product of remote sensing LE meets the requirements for use.