Accurate estimates of high-spatial-resolution global terrestrial latent heat flux (LE) from Landsat data are crucial for many basic and applied research. Yet current Landsat-derived LE products were developed using single algorithm with large uncertainties and discrepancies. Here we proposed a convolutional neural network-long short-term memory (CNN-LSTM)-based integrated LE (CNN-LSTM-ILE) framework that integrates five Landsat-derived physical LE algorithms, topography-related variables (elevation, slope and aspect) and eddy covariance (EC) observations to estimate 30-m global terrestrial LE. CNN-LSTM-ILE not only conserves good performance of LE estimation from pure deep learning (DL) algorithm, but partially inherits physical mechanism of the individual physical algorithms for improving the generalization of the integration algorithms for extreme cases. CNN-LSTM is an algorithm that combines two deep learning structures (CNN and LSTM) to effectively utilize the spatial and temporal information contained in the forcing inputs. The data were collected from 190 globally distributed EC observations from 2000 to 2015 and were provided by FLUXNET. The cross-validation results indicated that the CNN-LSTM integration algorithm improved the LE estimates by reducing the root mean square error (RMSE) of 5–8 W/m and increasing Kling-Gupta efficiency (KGE) of 0.05–0.16 when compared with the individual LE algorithms and the results of three other machine learning integration algorithms (multiple linear regression, MLR; random forest, RF; and deep neural networks, DNN). The CNN-LSTM integration algorithm had highest KGE (0.81) and R (0.66) compared to ground-measured and was applied to generate the Landsat-like regional and global terrestrial LE. An innovation of our strategy is that the CNN-LSTM-ILE model integrates pixel proximity effects and daily LE variations to enhance the accuracy of 16-day LE estimations. This approach can produce a more reliable Landsat-like global terrestrial LE product to improve the representativeness of heterogeneous regions for monitoring hydrological variables.

中文翻译：

---

使用广义深度 CNN-LSTM 集成算法对类 Landsat 全球陆地潜热通量进行多模型集成估计

利用 Landsat 数据准确估计高空间分辨率的全球陆地潜热通量 (LE) 对于许多基础和应用研究至关重要。然而，目前的陆地卫星衍生 LE 产品是使用单一算法开发的，具有很大的不确定性和差异。在这里，我们提出了一种基于卷积神经网络-长短期记忆（CNN-LSTM）的集成LE（CNN-LSTM-ILE）框架，该框架集成了五种源自Landsat的物理LE算法、地形相关变量（高程、坡度和坡向） ）和涡度协方差（EC）观测来估计 30 米全球陆地 LE。 CNN-LSTM-ILE不仅保留了纯深度学习（DL）算法的LE估计的良好性能，而且部分继承了各个物理算法的物理机制，以提高极端情况下集成算法的泛化能力。 CNN-LSTM 是一种结合了两种深度学习结构（CNN 和 LSTM）的算法，可有效利用强制输入中包含的空间和时间信息。这些数据是从 2000 年至 2015 年期间全球分布的 190 个 EC 观测中收集的，由 FLUXNET 提供。交叉验证结果表明，与传统算法相比，CNN-LSTM 集成算法将均方根误差 (RMSE) 降低了 5-8 W/m，并将 Kling-Gupta 效率 (KGE) 提高了 0.05-0.16，从而改善了 LE 估计。各个 LE 算法和其他三种机器学习集成算法（多元线性回归，MLR；随机森林，RF；和深度神经网络，DNN）的结果。与地面测量相比，CNN-LSTM 集成算法具有最高的 KGE (0.81) 和 R (0.66)，并用于生成类似 Landsat 的区域和全球陆地 LE。我们策略的一个创新是，CNN-LSTM-ILE 模型集成了像素邻近效应和每日 LE 变化，以提高 16 天 LE 估计的准确性。这种方法可以产生更可靠的类似 Landsat 的全球陆地 LE 产品，以提高监测水文变量的异质区域的代表性。