Estimation of crop evapotranspiration from MODIS data by combining random forest and trapezoidal models
Estimation of crop evapotranspiration from MODIS data by combining random forest and trapezoidal models
复制标题
DOI：
10.1016/j.agwat.2021.107249
复制
发表时间：
2022-01
影响因子：
6.7
通讯作者：
Pengyu Hao;L. Di;Liying Guo
中科院分区：
文献类型：
--
作者：
Pengyu Hao;L. Di;Liying Guo
蒸散量（ET）是作物生长监测和地表建模的重要参数。结合梯形模型和随机森林（RF）算法的优点，提出了一种利用MODIS数据计算作物生长季ET的新工作流程ESVEP-RF。在ESVEP-RF中，首先使用基于端元的土壤和植被能量划分（ESVEP）模型根据MODIS和气象输入计算一系列参数，然后将来自遥感数据、气象数据和ESVEP模型的所有参数用作RF算法的输入，进行潜热通量（LE）计算。利用位于内布拉斯加州（NE）和密歇根州（MI）的五个通量塔的12年（2003-2012年、2018年和2019年）的原位数据来测试ESVEP-RF的性能，结果表明，当训练样本数量充足且具有代表性时，ESVEP-RF在准确计算ET方面具有巨大潜力。 2010年和2011年，LE的R2约为0.8，RMSE约为70 W/m2，优于原始ESVEP模型结果。这表明RF算法能够更好地描述LST/FVC空间端元与LE之间的非线性相关性。在所有参数中，LAI、PLEv 和 R-vw 贡献较高，重要性百分比分别为 18.49%、15.71% 和 13.57%。此外，使用 2003 年至 2012 年从三个 NE 站点收集的所有样本来训练 RF 模型，然后计算 2018 年和 2019 年 NE 和 MI 站点的 LE。在 NE 站点中，RMSE 约为 65 W/m2，R2 约为 0.8。在 MI 站点中，值得注意的是，训练数据集中没有包含这些站点的样本，RMSE 约为 70 W/m2，R2 高于 0.7。这些结果显示了 ESVEP-RF 在提供最新 ET 信息方面的潜力。
Evapotranspiration (ET) is an important parameter for crop growth monitoring and land surface modeling. This paper proposed a new workflow, namely ESVEP-RF, to calculate ET during the crop growing season using MODIS data by combining the advantages of the trapezoidal model and Random Forest (RF) algorithm. In ESVEP-RF, the endmember-based soil and vegetation energy partitioning (ESVEP) model was first used to calculate a series of parameters from MODIS and meteorological inputs, and then all parameters derived from remote sensing data, meteorological data and ESVEP models were used as inputs to the RF algorithm for latent heat flux (LE) calculation. In-situ data of 12 years (2003–2012, 2018 and 2019) from five flux towers located in Nebraska (NE) and Michigan (MI) were used to test the performance of ESVEP-RF, and results showed that ESVEP-RF had great potential to accurately calculate ET when the number of training samples was sufficient and representative. In 2010 and 2011, R2of LE were around 0.8 and RMSE were around 70 W/m2, which outperformed original ESVEP model results. This indicated that the RF algorithm could better describe the non-linear correlation between in LST/FVC space endmembers and LE. Among all parameters, LAI, PLEv and R-vw had high contribution with percentage importance of 18.49%, 15.71% and 13.57%, respectively. Furthermore, all samples between 2003 and 2012 collected from the three NE sites were used to train RF models and then calculate LE for both NE and MI sites in 2018 and 2019. In NE sites, RMSE was around 65 W/m2and R2was around 0.8. In MI sites, it was noted that no samples from these sites were included in the training data set, and RMSE was around 70 W/m2and R2was higher than 0.7. These results showed the potential of ESVEP-RF for providing up-to-date ET information.