MODIS Evapotranspiration Downscaling Using a Deep Neural Network Trained Using Landsat 8 Reflectance and Temperature Data
MODIS Evapotranspiration Downscaling Using a Deep Neural Network Trained Using Landsat 8 Reflectance and Temperature Data
复制标题
DOI：
10.3390/rs14225876
复制
发表时间：
2022-11
期刊：
影响因子：
--
通讯作者：
Xianghong Che;Hankui K. Zhang;Qing Sun;Zutao Ouyang;Jiping Liu
中科院分区：
文献类型：
--
作者：
Xianghong Che;Hankui K. Zhang;Qing Sun;Zutao Ouyang;Jiping Liu
MODIS 8天综合蒸散(ET)产品(MOD16A2)被广泛用于研究大尺度的水文循环和能量收支。然而，MOD16A2的空间分辨率(500m)对于农业应用中的地方和区域水资源管理来说过于粗糙。在这项研究中，我们提出了一种基于深度神经网络(DNN)的MOD16A2降尺度方法，利用Landsat 8地表反射率和温度以及AgERA5气象变量来生成30m ET。该模型以MOD16A2 ET为参考，以500m分辨率进行训练，并应用于Landsat 8 30m分辨率。该方法在美国三个农业研究地点的15幅Landsat 8图像上进行了测试，并与通常用于ET缩减的经典随机森林回归模型进行了比较。应用于DNN回归模型的所有评价样本集的R2均高于随机森林模型(分别为0.64、2.76 mm/8d和14.92%)，而均方根偏差(RMSD)和相对RMSD(RRMSD)(平均值分别为0.67、2.63 mm/8d和14.92%)较低。与500米MOD16A2相比，DNN和随机森林缩小了30m的ET地图在空间上都有明显的改善，而DNN缩小的ET似乎更符合地表覆盖的变化。与原位ET测量结果(AmeriFlux)比较，DNN-Et具有更高的准确度，R2为0.73，RMSD为5.99 mm/8d，rRMSD为48.65%，而MOD16A2 ET为0.65，7.18和50.42%。
The MODIS 8-day composite evapotranspiration (ET) product (MOD16A2) is widely used to study large-scale hydrological cycle and energy budgets. However, the MOD16A2 spatial resolution (500 m) is too coarse for local and regional water resource management in agricultural applications. In this study, we propose a Deep Neural Network (DNN)-based MOD16A2 downscaling approach to generate 30 m ET using Landsat 8 surface reflectance and temperature and AgERA5 meteorological variables. The model was trained at a 500 m resolution using the MOD16A2 ET as reference and applied to the Landsat 8 30 m resolution. The approach was tested on 15 Landsat 8 images over three agricultural study sites in the United States and compared with the classical random forest regression model that has been often used for ET downscaling. All evaluation sample sets applied to the DNN regression model had higher R2 and lower root-mean-square deviations (RMSD) and relative RMSD (rRMSD) (the average values: 0.67, 2.63 mm/8d and 14.25%, respectively) than the random forest model (0.64, 2.76 mm/8d and 14.92%, respectively). Spatial improvement was visually evident both in the DNN and the random forest downscaled 30 m ET maps compared with the 500 m MOD16A2, while the DNN-downscaled ET appeared more consistent with land surface cover variations. Comparison with the in situ ET measurements (AmeriFlux) showed that the DNN-downscaled ET had better accuracy, with R2 of 0.73, RMSD of 5.99 mm/8d and rRMSD of 48.65%, than the MOD16A2 ET (0.65, 7.18 and 50.42%, respectively).