Spatial Predictions of Evapotranspiration from Tropical Landscapes Using Remote Sensing Data
Doctoral thesis
Date of Examination:2025-08-21
Date of issue:2025-12-11
Advisor:Prof. Dr. Dirk Hoelscher
Referee:Prof. Dr. Dirk Hoelscher
Referee:Prof. Dr. Alexander Knohl
Referee:Prof. Dr. Dominik Seidel
Files in this item
Name:PhD Thesis.pdf
Size:27.8Mb
Format:PDF
Description:PhD Thesis_Cummulative Publications_Alejandra Valdes Uribe
Abstract
English
Tropical forests play a crucial role in global terrestrial evapotranspiration (ET). As these forests continue to be converted to other land uses, the resulting changes in ET could have significant implications for both regional and global climate systems. However, many questions about forest ET remain open, particularly in drylands with varying aridity, mountainous areas, and forest fragments that create variable conditions across the space. Recent advances in remote sensing have opened new possibilities for studying ET in the tropics. Key developments include improvements in spaceborne thermal imaging capabilities, LiDAR for vegetation structure assessments, and the availability of open-source datasets. These, combined with advanced spatial analysis methods, provide unprecedented opportunities for studying ET in the tropics. Building on these developments, the overall objective of this dissertation is to advance understanding of how critical biophysical factors, including vegetation structure, topographic features, climatic conditions, and landscape configuration, influence spatial ET variation across tropical forest landscapes. Study 1 addressed ET across a dryland ecotone in western Ecuador, where the spatial distribution of vegetation characteristics is shaped by environmental conditions, with aridity being a key factor. We hypothesized that vegetation structural features significantly contribute to the spatial variation in ET during both wet and dry seasons. We assessed three-dimensional vegetation structure using a ground-based mobile laser scanner on 75 plots representing different forest types. ET data were retrieved from MODIS satellite observations. During the wet season, models incorporating spatial autocorrelation through Moran’s Eigenvector Maps explained 35% and 36% of spatial ET variability using plant area index and structural complexity (Db), respectively. In the dry season, same predictors explained over 70% of the spatial variation in ET. Our results demonstrate that differences in vegetation structure, particularly Db, significantly contribute to explaining spatial ET variation, with this relationship being especially pronounced during the dry season. We propose that dry-season aridity drives structural adaptations in vegetation that reduce ecosystem ET rates. Our findings emphasize the importance of preserving the complex structural diversity of dryland ecosystems to ensure their continued functionality. Study 2 investigated ET patterns across a large, protected tropical forest region on the western slope of the Andes. Our objectives were to identify the combination of variables most relevant for spatial ET prediction across the region and to determine the specific role of forest structure in spatial ET modelling. We employed a random forest (RF) approach following established best practices for spatial predictions to address spatial autocorrelation. The analysis utilized topographic, meteorological, and forest structure variables derived from open-source products including GEDI, PROBA-V, and ERA5, excluding any variables incorporated in the ECOSTRESS-L3 algorithm to ensure model independence. The models demonstrated high accuracy for spatially explicit ET prediction across different locations, achieving R2 values ranging from 0.61 to 0.74. Notably, no single predictor dominated the models, with five key variables collectively explaining 60% of model performance. Leaf area index emerged as one of the three most influential variables for ET prediction across all study days, alongside the topographic variables of elevation and aspect. We conclude that the random forest approach provides robust ET predictions that could help address observation gaps in tropical regions. Our findings highlight the importance of considering both terrain characteristics and vegetation properties in understanding tropical forest water cycles. Study 3 examined forest fragments within human-dominated landscapes across the lowlands and Andes of western Ecuador. We analysed the spatial pattern of ET from 83 forest fragments using spaceborne ECOSTRESS data. Our objective was to investigate the extent and direction of edge effects on ET, focusing on changes relative to interior forest conditions, while considering how abiotic and biotic factors interact to contribute to these observed variations at forest edges. We applied changepoint analysis and found that the frequency of ET shifts increased progressively from the interior of fragments toward the edge, where ET was enhanced in most cases (61%). A RF model with target-oriented crossvalidation achieved a spatial prediction accuracy of 64%, identifying elevation, aridity, and distance to edge as the most important predictors. SHAP analysis also highlighted the importance of edge effects and indicated greater enhancement of ET, particularly at mid-elevations and in areas with high canopy cover. These patterns likely resulted from a combination of factors, including climatic conditions, forest structure, edge orientation, and surrounding land-use types. Our findings suggest that forest fragments may enhance ET through edge effects, though the high variability observed at edges indicates need for further investigation. While conserving large old-growth forests remains the primary strategy for maintaining climate regulation functions in tropical landscapes, our study indicates that the numerous small forest fragments—which have been disappearing rapidly in recent decades—may also play an important role in regional water cycles. Study 4 investigated a heterogeneous tropical landscape in northeastern Madagascar, encompassing continuous old-growth forest, forest fragments, agricultural land, agroforests, and recovering forests. Our research had three primary objectives: to identify combinations of meteorological, topographical, soil, and vegetation structure variables most relevant for spatially predicting ET in the region; to address observation gaps in ET data; and to quantify the relative contributions and interactions of vegetation structure variables in spatial prediction models. Daily ET data were obtained from ECOSTRESS, while predictor variables were sourced from multiple open datasets. To address potential spatial autocorrelation, we implemented forward feature selection combined with target-oriented cross-validation. RF models demonstrated high predictive accuracy for spatial ET estimation, achieving R2 values ranging from 0.7 to 0.9 across different days. Using SHAP for model interpretation, we found that meteorological variables contributed most substantially to predictions (51 - 62%), followed by vegetation structure (15 - 18%), topographical features (6 - 24%), and soil properties (8 - 13%). The SHAP analysis further revealed critical interactions between variables, particularly demonstrating how vegetation structure influences ET responses under varying rainfall and wind speed conditions. In conclusion, the four studies show that the spatial variations of ET in tropical landscapes can be explained from the interplay between vegetation structure, topography, meteorology, and spatial configuration. Forest structural characteristics repeatedly ranked among the strongest predictors, and showed interactions with topography and climate. Forest edges further modulated ET patterns, highlighting the potential role of forest fragments in the water cycle.
Keywords: ECOSTRESS; Vegetation structural complexity; random forest; spatial autocorrelation; SHAP analysis; Ecuador
