Abstract. Evapotranspiration (ET) quantification improves the comprehension of the water, heat, and carbon interactions and the feedback to the climate, which is essential for global change research. We aimed to model ET using artificial neural networks (ANNs) based on Landsat-8 and reanalysis data from the National Centers for Environmental Prediction over the grasslands of the Pampa biome. The output variable was the ET trained by eddy covariance (EC) measurements acquired from a flux tower located in Santa Maria, Brazil. ANN was performed using the backpropagation algorithm with four remote sensing input variables (albedo, normalized difference vegetation index, land surface temperature, and surface net radiation). In addition, four meteorological variables from the Environmental Prediction Climate Forecast System Version 2 hourly product were included in the model (air temperature, atmospheric pressure, relative humid, and wind speed). We analyzed 67 clear-sky scenes between 2014 and 2019. Results produced very robust daily ET estimates. ANN exhibited a correlation of 0.88 relative to in situ EC data, demonstrating a good linear relationship between ET estimated and measured and producing a root-mean-square error (mean absolute error) of 0.75 (0.58) mm/day. The ANN model was also compared with the widely known simplified surface energy balance index (S-SEBI) model. S-SEBI exhibited lower correlation with the ET in situ compared to the ANN model. Furthermore, the ANN model had a superior performance in summer and winter seasons in which S-SEBI was found to outperform the ET in situ. The model developed in our research is an alternative to approaches that need a great number of input variables or in situ data since it is only dependent on freely available data. Therefore, it should support future integrated strategies of water resources allocation over the natural grasslands of the Brazilian Pampa.

中文翻译：

---

基于遥感的人工神经网络模型反演巴西潘帕草原的蒸散量

摘要。蒸散 (ET) 量化提高了对水、热和碳相互作用以及对气候反馈的理解，这对于全球变化研究至关重要。我们的目标是使用基于 Landsat-8 的人工神经网络 (ANN) 和来自国家环境预测中心对潘帕生物群落草原的再分析数据对 ET 进行建模。输出变量是通过从位于巴西圣玛丽亚的通量塔获取的涡度协方差 (EC) 测量值训练的 ET。ANN 使用反向传播算法与四个遥感输入变量（反照率、归一化差异植被指数、地表温度和地表净辐射）执行。此外，模型中包含来自环境预测气候预测系统第 2 版每小时产品的四个气象变量（气温、大气压力、相对湿度和风速）。我们分析了 2014 年至 2019 年间的 67 个晴空场景。结果产生了非常可靠的每日 ET 估计值。ANN 与原位 EC 数据的相关性为 0.88，表明 ET 估计和测量之间存在良好的线性关系，并产生 0.75 (0.58) mm/天的均方根误差（平均绝对误差）。ANN 模型还与广为人知的简化表面能平衡指数 (S-SEBI) 模型进行了比较。与 ANN 模型相比，S-SEBI 与 ET 的相关性较低。此外，ANN 模型在夏季和冬季具有优越的性能，其中 S-SEBI 被发现优于 ET 原位。我们研究中开发的模型是需要大量输入变量或原位数据的方法的替代方法，因为它仅依赖于免费可用的数据。因此，它应该支持未来巴西潘帕草原天然草原水资源配置的综合战略。