As climate change progresses, urban areas are increasingly affected by water scarcity and the urban heat island effect. Evapotranspiration (ET) is a crucial component of urban greening initiatives of cities worldwide aimed at mitigating these issues. However, ET estimation methods in urban areas have so far been limited. An expanding number of flux towers in urban environments provide the opportunity to directly measure ET by the eddy covariance method. In this study, we present a novel approach to model urban ET by combining flux footprint modeling, remote sensing and geographic information system (GIS) data, and deep learning and machine learning techniques. This approach facilitates spatio-temporal extrapolation of ET at a half-hourly resolution; we tested this approach with a two-year dataset from two flux towers in Berlin, Germany. The benefit of integrating remote sensing and GIS data into models was investigated by testing four predictor scenarios. Two algorithms (1D convolutional neural networks (CNNs) and random forest (RF)) were compared. The best-performing models were then used to model ET values for the year 2019. The inclusion of GIS data extracted using flux footprints enhanced the predictive accuracy of models, particularly when meteorological data was more limited. The best-performing scenario (meteorological and GIS data) showed an RMSE of 0.0239 mm/h and R2 of 0.840 with RF and an RMSE of 0.0250 mm/h and a R2 of 0.824 with 1D CNN for the more vegetated site. The 2019 ET sum was substantially higher at the site surrounded by more urban greenery (366 mm) than at the inner-city site (223 mm), demonstrating the substantial influence of vegetation on the urban water cycle. The proposed method is highly promising for modeling ET in a heterogeneous urban environment and can support climate change mitigation initiatives of urban areas worldwide.

The monitoring of soil moisture content (SMC) at very high spatial resolution (<10m) using unmanned aerial systems (UAS) is of high interest for precision agriculture and the validation of large scale SMC products. Data-driven approaches are the most common method to retrieve SMC with UAS-borne data at water limited sites over non-disturbed agricultural crops. A major disadvantage of data-driven algorithms is the limited transferability in space and time and the need of a high number of ground reference samples. Physically-based approaches are less dependent on the amount of samples and are transferable in space and time. This study explores the potential of (1) a hybrid method targeting the soil brightness factor of the PROSAIL model using a variational heteroscedastic Gaussian Processes regression (VHGPR) algorithm, and (2) a data-driven method employing VHGPR for the retrieval of SMC over three grassland sites based on UAS-borne VIS-NIR (399-1001 nm) hyperspectral data. The sites were managed by mowing (Fendt), grazing (Grosses Bruch) and irrigation (Marquardt). With these distinct local pre-conditions we aimed to identify factors that favor and limit the retrieval of SMC.

Spectral, temporal and spatial dimensions are difficult to model together when predicting in situ plant traits from remote sensing data. Therefore, machine learning algorithms solely based on spectral dimensions are often used as predictors, even when there is a strong effect of spatial or temporal autocorrelation in the data. A significant reduction in prediction accuracy is expected when algorithms are trained using a sequence in space or time that is unlikely to be observed again. The ensuing inability to generalise creates a necessity for ground-truth data for every new area or period, provoking the propagation of “single-use” models. This study assesses the impact of spatial autocorrelation on the generalisation of plant trait models predicted with hyperspectral data. Leaf Area Index (LAI) data generated at increasing levels of spatial dependency are used to simulate hyperspectral data using Radiative Transfer Models. Machine learning regressions to predict LAI at different levels of spatial dependency are then tuned (determining the optimum model complexity) using cross-validation as well as the NOIS method. The results show that cross-validated prediction accuracy tends to be overestimated when spatial structures present in the training data are fitted (or learned) by the model.