Environmental changes in biodiversity hotspot ecosystems of
South Ecuador: RESPonse and feedback effECTs (FOR2730)

Abstract:

A new satellite-based algorithm for rainfall retrieval in high spatio-temporal resolution fo Ecuador is presented. The algorithm relies on the precipitation information from the Integrated Multi-SatEllite Retrieval for the Global Precipitation Measurement (GPM) (IMERG) and infrared (IR) data from the Geostationary Operational Environmental Satellite-16 (GOES-16). It wa developed to (i) classify the rainfall area (ii) assign the rainfall rate. In each step, we selected the most important predictors and hyperparameter tuning parameters monthly. Between 19 April 2017 and 30 November 2017, brightness temperature derived from the GOES-16 IR channels and ancillary geo-information were trained with microwave-only IMERG-V06 using random forest (RF). Validation was done against independent microwave-only IMERG-V06 information not used for training. The validation results showed the new rainfall retrieva technique (multispectral) outperforms the IR-only IMERG rainfall product. This offers using the multispectral IR data can improve the retrieval performance compared to single-spectrum IR approaches. The standard verification scored a median Heidke skill score of ~0.6 for the rain area delineation and R between ~0.5 and ~0.62 for the rainfall rate assignment, indicating uncertainties for Andes’s high elevation. Comparison of RF rainfall rates in 2 km2 resolution with daily rain gauge measurements reveals the correlation of R = ~0.33.

Abstract:

The Random Forest (RF) algorithm, a decision-tree-based technique, has become a promising approach for applications addressing runoff forecasting in remote areas. This machine learning approach can overcome the limitations of scarce spatio-temporal data and physical parameters needed for process-based hydrological models. However, the influence of RF hyperparameters is still uncertain and needs to be explored. Therefore, the aim of this study is to analyze the sensitivity of RF runoff forecasting models of varying lead time to the hyperparameters of the algorithm. For this, models were trained by using (a) default and (b) extensive hyperparameter combinations through a grid-search approach that allow reaching the optimal set. Model performances were assessed based on the R2, %Bias, and RMSE metrics. We found that: (i) The most influencing hyperparameter is the number of trees in the forest, however the combination of the depth of the tree and the number of features hyperparameters produced the highest variability-instability on the models. (ii) Hyperparameter optimization significantly improved model performance for higher lead times (12- and 24-h). For instance, the performance of the 12-h forecasting model under default RF hyperparameters improved to R2 = 0.41 after optimization (gain of 0.17). However, for short lead times (4-h) there was no significant model improvement (0.69 < R2 < 0.70). (iii) There is a range of values for each hyperparameter in which the performance of the model is not significantly affected but remains close to the optimal. Thus, a compromise between hyperparameter interactions (i.e., their values) can produce similar high model performances. Model improvements after optimization can be explained from a hydrological point of view, the generalization ability for lead times larger than the concentration time of the catchment tend to rely more on hyperparameterization than in what they can learn from the input data. This insight can help in the development of operational early warning systems.

Abstract:

Despite many eorts of the radar community, quantitative precipitation estimation (QPE) from weather radar data remains a challenging topic. The high resolution of X-band radar imagery in space and time comes with an intricate correction process of reflectivity. The steep and high mountain topography of the Andes enhances its complexity. This study aims to optimize the rainfall derivation of the highest X-band radar in the world (4450 m a.s.l.) by using a random forest (RF) model and single Plan Position Indicator (PPI) scans. The performance of the RFmodel was evaluated in comparison with the traditional step-wise approach by using both, the Marshall-Palmer and a site-specific Z–R relationship. Since rain gauge networks are frequently unevenly distributed and hardly available at real time in mountain regions, bias adjustment was neglected. Results showed an improvement in the step-wise approach by using the site-specific (instead of the Marshall-Palmer) Z–R relationship. However, both models highly underestimate the rainfall rate (correlation coecient < 0.69; slope up to 12). Contrary, the RF model greatly outperformed the step-wise approach in all testing locations and on dierent rainfall events (correlation coecient up to 0.83; slope = 1.04). The results are promising and unveil a dierent approach to overcome the high attenuation issues inherent to X-band radars.

Abstract:

Soils of Páramo ecosystems regulate the water supply to many Andean populations. In spite of being a necessary input to distributed hydrological models, regionalized soil water retention data from these areas are currently not available. The investigated catchment of the Quinuas River has a size of about 90 km2 and comprises parts of the Cajas National Park in southern Ecuador. It is dominated by soils with high organic carbon contents, which display characteristics of volcanic influence. Besides providing spatial predictions of soil water retention at the catchment scale, the study presents a detailed methodological insight to model setup and validation of the underlying machine learning approach with random forest. The developed models performed well predicting volumetric water contents between 0.55 and 0.9 cm3 cm? 3. Among the predictors derived from a digital elevation model and a Landsat image, altitude and several vegetation indices provided the most information content. The regionalized maps show particularly low water retention values in the lower Quinuas valley, which go along with high prediction uncertainties. Due to the small size of the dataset, mineral soils could not be separated from organic soils, leading to a high prediction uncertainty in the lower part of the valley, where the soils are influenced by anthropogenic land use.

Abstract:

Landslides are a hazard for humans and artificial structures. From an ecological point of view, they represent an important ecosystem disturbance, especially in tropical montane forests. Here, shallow translational landslides are a frequent natural phenomenon and one local determinant of high levels of biodiversity. In this paper, we apply weighted ensembles of advanced phenomenological models from statistics and machine learning to analyze the driving factors of natural landslides in a tropical montane forest in South Ecuador. We exclusively interpret terrain attributes, derived from a digital elevation model, as proxies to several driving factors of landslides and use them as predictors in our models which are trained on a set of five historical landslide inventories. We check the model generality by transferring them in time and use three common performance criteria (i.e. AUC, explained deviance and slope of model calibration curve) to, on the one hand, compare several state-of-the-art model approaches and on the other hand, to create weighted model ensembles. Our results suggest that it is important to consider more than one single performance criterion. Approaching our main question, we compare responses of weighted model ensembles that were trained on distinct functional units of landslides (i.e. initiation, transport and deposition zones). This way, we are able to show that it is quite possible to deduce driving factors of landslides, if the consistency between the training data and the processes is maintained. Opening the ?black box? of statistical models by interpreting univariate model response curves and relative importance of single predictors regarding their plausibility, we provide a means to verify this consistency. With the exception of classification tree analysis, all techniques performed comparably well in our case study while being outperformed by weighted model ensembles. Univariate response curves of models trained on distinct functional units of landslides exposed different shapes following our expectations. Our results indicate the occurrence of landslides to be mainly controlled by factors related to the general position along a slope (i.e. ridge, open slope or valley) while landslide initiation seems to be favored by small scale convexities on otherwise plain open slopes.