ECOSTRESS Reveals the Importance of Topography and Forest Structure for Evapotranspiration from a Tropical Forest Region of the Andes

Abstract

Tropical forests are major sources of global terrestrial evapotranspiration (ET), but these heterogeneous landscapes pose a challenge for continuous estimates of ET, so few studies are conducted, and observation gaps persist. New spaceborne products such as ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) are promising tools for closing such observation gaps in understudied tropical areas. Using ECOSTRESS ET data across a large, protected tropical forest region (2250 km2) situated on the western slope of the Andes, we predicted ET for different days. ET was modeled using a random forest approach, following best practice workflows for spatial predictions. We used a set of topographic, meteorological, and forest structure variables from open-source products such as GEDI, PROBA-V, and ERA5, thereby avoiding any variables included in the ECOSTRESS L3 algorithm. The models indicated a high level of accuracy in the spatially explicit prediction of ET across different locations, with an r2 of 0.61 to 0.74. Across all models, no single predictor was dominant, and five variables explained 60% of the models’ results, thus highlighting the complex relationships among predictor variables and their influence on ET spatial predictions in tropical mountain forests. The leaf area index, a forest structure variable, was among the three variables with the highest individual contributions to the prediction of ET on all days studied, along with the topographic variables of elevation and aspect. We conclude that ET can be predicted well with a random forest approach, which could potentially contribute to closing the observation gaps in tropical regions, and that a combination of topography and forest structure variables plays a key role in predicting ET in a forest on the western slope of the Andes.

UAV-based thermography reveals spatial and temporal variability of evapotranspiration from a tropical rainforest

“…The trees at the riparian sites were of smaller stature, had on average 43% lower biomass compared to the upland sites and showed signs of forest disturbance (Ahongshangbam et al, 2020;Kotowska and Waite, unpublished results;Rembold et al, unpublished results). There is evidence from studies correlating ET rate to forest structure variables such as leaf area index (LAI), sapwood area, height and tree biomass with ET in general increasing with these variables (Álvarez-Dávila et al, 2017;Jaramillo et al, 2018;Metzen et al, 2019;Valdés-Uribe et al, 2023). The history of disturbance and the generally smaller trees at the riparian sites suggests regenerating forest stands, which may also influence transpiration rates (Ghimire et al, 2022).…”

Combining UAV thermography, point cloud analysis and machine learning for assessing small‐scale evapotranspiration patterns in a tropical rainforest

“…For spatial predictions in ecological studies, random forest stands out among the available algorithms as particularly well performing (Ahmad et al, 2017;Fernandez-Delgado et al, 2014), for example, when applied to predict reference ET (Dias et al, 2021;Feng et al, 2017), ET of tropical mountain forests (Valdés-Uribe et al, 2023), water stress (Virnodkar et al, 2020), sap flux and leaf stomatal conductance (Ellsäßer, Röll, Ahongshangbam, et al, 2020), net ecosystem exchange (Reitz et al, 2021) or land-cover change (Aide et al, 2013). Recent studies have proposed solutions to previous autocorrelation and overfitting issues in spatial predictions via forward feature selection (FFS) and target oriented cross validation, thus minimizing the risk of spatial overfitting with random forests and showing realistic overall model performances (Gasch et al, 2015;Meyer et al, 2018Meyer et al, , 2019.…”

“…To our knowledge, there are no studies available for comparison involving such a set of structural canopy variables for predicting ET under given climatic conditions at such small spatial scales. A random forest study predicting ET at much larger spatial scales (70 m pixel size) found a dominant influence of topographic variables, and to a lesser extent of forest structure variables such as the leaf area index, on spatial predictions of ET(Valdés-Uribe et al, 2023); while climatic variables were found to be of relatively minor importance in that study, some further previous studies, also at much larger spatial scales, successfully applied machine learning techniques to model ET (dynamics) from climatic datasets(Granata, 2019;Tikhamarine et al, 2019). Both the random forest ET study by Valdés-Uribe et al (2023) and our study show varying variable importance across different studied days in tropical forests, which may (partially) be related to differences in the given climatic conditions.…”