Moisture cycling responses of an undisturbed tropical woodland to climate variability using eddy covariance and machine learning

Abstract:

Higher energy availability and unique biodiversity on the tropics require field hydrology studies, where undisturbed environments may perform hydrological functions that need to be well understood. We aimed to assess the water partitioning behavior and variability in the critical zone of a wooded Cerrado fragment (a tropical woodland). Thus, we performed flux measurements over the area located in Southeastern Brazil using eddy covariance; used machine learning techniques to obtain a model to estimate the evapotranspiration (ET) of the wooded Cerrado area using meteorological data; and simulated a long term water balance using a stochastic climate generator inputs and the previously calibrated ET model. The average ET for the wooded Cerrado along 253 days was 3.12 ± 0.93 mm d^-1, and the uncertainties were close to ± 20%. We trained models to estimate ET using machine learning using the observed ET (response) and meteorological data (explanatory variables): solar radiation (Rg), wind speed (WS), temperature (T), relative humidity (RH), and rainfall (P). We found that the k-nearest neighbors (KNN) presented the best performance, when compared to other techniques: decision trees (DTR), neural networks multilayer perceptron (MLP), and ensemble based AdaBoost (ADA). We could simulate the water balance by using a stochastic weather generator, whose outcomes were obtained on a daily basis: Rg, WS, T, RH, and P. These data were used to calculate long-term ET and overland flow (OF), with the previously trained ET model and a literature runoff coefficient, respectively. Hence, we could calculate the water balance residuals (dS/dt) by subtracting ET and OF from P. The variability of the water cycle was assessed using the simulated data. The area presented a resilient behavior to droughts, showing similar periods of water scarcity and surplus, highlighting its importance for water storage in the subsurface and moisture cycling continuity along the whole year. Finally, the water balance at the long term relies on the rainfall stochasticity. This dependency can be observed on the average simulated dS/dt (172 ± 211 mm yr^-1) and P (1227 ± 208 mm yr^-1), which have similar standard deviations (around 200 mm yr^-1); differently from the ET (1054 ± 46 mm yr^-1), which presented high annual rates but a small variability along the simulated years.