Dokument

Summary:
Rainfall in the tropical Andes high mountains is paramount for understanding complex hydrological and ecological phenomena that take place in this distinctive area of the world. Here, rainfall drives imminent hazards such as severe floods, rainfall-induced landslides, different types of erosion, among others. Nonetheless, sparse and uneven distributed rain gauge networks as well as low- resolution satellite imagery are not sufficient to capture its high variability and complex dynamics in the irregular topography of high mountains at appropriate temporal and spatial scales. This results in both, a lack of knowledge about rainfall patterns, as well as a poor understanding of rainfall microphysics, which to date are largely underexplored in the tropical Andes. Therefore, this investigation focuses on the deployment and exploitation of single-polarization (SP) X-band weather radars in the Andean high mountain regions of southern Ecuador, applicable to quantitative precipitation estimation (QPE) and discharge forecasting. This work leverages radar rainfall data by exploring a machine learning (ML) approach. The main aims of the thesis were: (i) The deployment of a first X-band weather radar network in tropical high mountains, (ii) the physically-based QPE of X-band radar retrievals, (iii) the optimization of radar QPE by using a ML-based model and (iv) a discharge forecasting application using a ML-based model and SP X-band radar data. As a starting point, deployment of the first weather radar network in tropical high mountains was carried out. A complete framework for data transmission was set for communication among the network. The highest radar in the network (4450 m a.s.l.) was selected in this study for exploiting the potential of SP X-band radar data in the Andes. First and foremost, physically-based QPE was performed through the derivation of Z-R relationships. For this, data from three disdrometers at different geographic locations and elevation were used. Several rainfall events were selected in order to perform a classification of rainfall types based on the mean volume diameter (Dm [mm]). Derived Z-R relations confirmed the high variability in their parameters due to different rainfall types in the study area. Afterwards, the optimization of radar QPE was pursued by using a ML approach as an alternative to the common physically-based QPE method by means of the Z-R relation. For this, radar QPE was tackled by using two different approaches. The first one was conducted by implementing a step-wise approach where reflectivity correction is performed in a step-by-step basis (i.e., clutter removal, attenuation correction). Finally a locally derived Z-R relationship was applied for obtaining radar QPE. Rain gauge-bias adjustment was neglected because the availability of rain gauge data at near-real time is limited and infrequent in the study area. The second one was conducted by an implementation of a radar QPE model that used the Random Forest (RF) algorithm and reflectivity derived features as inputs for the model. Finally, the performances of both models were compared against rain gauge data. The results showed that the ML-based model outperformed the step-wise approach, making it possible to obtain radar QPE without the need of rain gauge data after the model was implemented. It also allowed to extend the useful range of the radar image (i.e., up to 50 km). Radar QPE can be generally used as input for discharge forecasting models if available. However, one could expect from ML-based models as RF, the ability to map radar data to the target variable (discharge) without any intermediate step (e.g., transformation from reflectivity to rainfall rate). Thus, a comparison for discharge forecasting was performed between RF models that used different input data type. Input data for the relevant models were obtained either from native reflectivity records (i.e., reflectivity corrected from unrealistic measurements) or derived radar-rainfall data (i.e., radar QPE). Results showed that both models performed alike. This proved the suitability of using native radar data (reflectivity) for discharge forecasting in mountain regions. This could be extrapolated in the advantages of deploying radar networks and use their information directly to fed early-warning systems regardless of the availability of rain gauges at ground. In summary, this investigation (i) participated on the deployment of the first weather radar network in tropical high mountains, (ii) significantly contributed to a deeper understanding of rainfall microphysics and its variability in the high tropical Andes by using disdrometer data and (iii) exploited, for the very first time, the native X-band radar reflectivity as a suitable input for ML-based models for both, optimized radar QPE and discharge forecasting. The latter highlighted the benefits and potentials of using a ML approach in radar hydrology. The research generally accounted for ground monitoring limitations commonly found in mountain regions and provided a promising alternative with leveraging the cost-effective X-band technology in the steep terrain of the Andean Cordillera.

Das Dokument ist im Internet frei zugänglich - Hinweise zu den Nutzungsrechten