abstract: Recently, a well-designed and well-trained neural network can yield state-of-the-art results across many domains, including data mining, computer vision, and medical image analysis. But progress has been limited for tasks where labels are difficult or impossible to obtain. This reliance on exhaustive labeling is a critical limitation in the rapid deployment of neural networks. Besides, the current research scales poorly to a large number of unseen concepts and is passively spoon-fed with data and supervision. To overcome the above data scarcity and generalization issues, in my dissertation, I first propose two unsupervised conventional machine learning algorithms, hyperbolic stochastic coding, and multi-resemble multi-target low-rank coding, to solve the incomplete data and missing label problem. I further introduce a deep multi-domain adaptation network to leverage the power of deep learning by transferring the rich knowledge from a large-amount labeled source dataset. I also invent a novel time-sequence dynamically hierarchical network that adaptively simplifies the network to cope with the scarce data. To learn a large number of unseen concepts, lifelong machine learning enjoys many advantages, including abstracting knowledge from prior learning and using the experience to help future learning, regardless of how much data is currently available. Incorporating this capability and making it versatile, I propose deep multi-task weight consolidation to accumulate knowledge continuously and significantly reduce data requirements in a variety of domains. Inspired by the recent breakthroughs in automatically learning suitable neural network architectures (AutoML), I develop a nonexpansive AutoML framework to train an online model without the abundance of labeled data. This work automatically expands the network to increase model capability when necessary, then compresses the model to maintain the model efficiency. In my current ongoing work, I propose an alternative method of supervised learning that does not require direct labels. This could utilize various supervision from an image/object as a target value for supervising the target tasks without labels, and it turns out to be surprisingly effective. The proposed method only requires few-shot labeled data to train, and can self-supervised learn the information it needs and generalize to datasets not seen during training. / Dissertation/Thesis / Doctoral Dissertation Computer Science 2020

Streamflow dynamics in mountainous environments are controlled by runoff generation processes in the basin upstream. Runoff generation processes are thus a major control of the terrestrial part of the water cycle, influencing both, water quality and water quantity as well as their dynamics. The understanding of these processes becomes especially important for the prediction of floods, erosion, and dangerous mass movements, in particular as hydrological systems often show threshold behavior. In case of extensive environmental changes, be it in climate or in landuse, the understanding of runoff generation processes will allow us to better anticipate the consequences and can thus lead to a more responsible management of resources as well as risks. In this study the runoff generation processes in a small undisturbed catchment in the Chilean Andes were investigated. The research area is characterized by steep hillslopes, volcanic ash soils, undisturbed old growth forest and high rainfall amounts. The investigation of runoff generation processes in this data scarce area is of special interest as a) little is known on the hydrological functioning of the young volcanic ash soils, which are characterized by extremely high porosities and hydraulic conductivities, b) no process studies have been carried out in this area at either slope or catchment scale, and c) understanding the hydrological processes in undisturbed catchments will provide a basis to improve our understanding of disturbed systems, the shift in processes that followed the disturbance and maybe also future process evolution necessary for the achievement of a new steady state. The here studied catchment has thus the potential to serve as a reference catchment for future investigations. As no long term data of rainfall and runoff exists, it was necessary to replace long time series of data with a multitude of experimental methods, using the so called "multi-method approach". These methods cover as many aspects of runoff generation as possible and include not only the measurement of time series such as discharge, rainfall, soil water dynamics and groundwater dynamics, but also various short term measurements and experiments such as determination of throughfall amounts and variability, water chemistry, soil physical parameters, soil mineralogy, geo-electrical soundings and tracer techniques. Assembling the results like pieces of a puzzle produces a maybe not complete but nevertheless useful picture of the dynamic ensemble of runoff generation processes in this catchment. The employed methods were then evaluated for their usefulness vs. expenditures (labour and financial costs). Finally, the hypotheses - the perceptual model of runoff generation generated from the experimental findings - were tested with the physically based model Catflow. Additionally the process-based model Wasim-ETH was used to investigate the influence of landuse on runoff generation at the catchment scale. An initial assessment of hydrologic response of the catchment was achieved with a linear statistical model for the prediction of event runoff coefficients. The parameters identified as best predictors give a first indication of important processes. Various results acquired with the "multi-method approach" show that response to rainfall is generally fast. Preferential vertical flow is of major importance and is reinforced by hydrophobicity during the summer months. Rapid lateral water transport is necessary to produce the fast response signal, however, while lateral subsurface flow was observed at several soil moisture profiles, the location and type of structures causing fast lateral flow on the hillslope scale is still not clear and needs to be investigated in more detail. Surface runoff has not been observed and is unlikely due to the high hydraulic conductivities of the volcanic ash soils. Additionally, a large subsurface storage retains most of the incident rainfall amount during events (>90%, often even >95%) and produces streamflow even after several weeks of drought. Several findings suggest a shift in processes from summer to winter causing changes in flow patterns, changes in response of stream chemistry to rainfall events and also in groundwater-surface water interactions. The results of the modelling study confirm the importance of rapid and preferential flow processes. However, due to the limited knowledge on subsurface structures the model still does not fully capture runoff response. Investigating the importance of landuse on runoff generation showed that while peak runoff generally increased with deforested area, the location of these areas also had an effect. Overall, the "multi-method approach" of replacing long time series with a multitude of experimental methods was successful in the identification of dominant hydrological processes and thus proved its applicability for data scarce catchments under the constraint of limited resources. / Die Abflussdynamik in Mittel- und Hochgebirgen wird durch die Abflussbildungsprozesse im Einzugsgebiet bestimmt. Diese Prozesse kontrollieren damit zu großen Teilen den terrestrischen Teil des Wasserkreislaufs und beeinflussen sowohl Wasserqualität als auch -quantität. Das Verständnis von Abflussbildungsprozessen ist besonders wichtig für die Vorhersage von Hochwasser, Erosion und Massenbewegungen (z.B. Erdrutsche) da hydrologische Systeme oft Schwellenwertverhalten aufweisen. Im Falle weit reichender Umweltveränderungen, wie z.B. Klima- oder Landnutzungsänderungen kann das Verständnis der Abflussbildungsprozesse ein verantwortungsvolleres Management sowohl der Ressourcen als auch der Risiken ermöglichen. In dieser Studie wurden die Abflussbildungsprozesse in einem kleinen, anthropogen unbeeinflussten Einzugsgebiet in den Chilenischen Anden untersucht. Das Untersuchungsgebiet ist durch steile Hänge, vulkanische Ascheböden, ungestörten Naturwald und hohe Niederschlagsmengen charakterisiert. Die Erforschung von Abflussbildungsprozessen ist hier von besonderem Interesse, da a) wenig über das hydrologische Verhalten der hochporösen und hochleitfähigen jungen Ascheböden bekannt ist, b) in dieser Region bisher keine Studien auf Hang- oder Einzugsgebietsskala durchgeführt wurden, und c) das Prozessverständnis in ungestörten Einzugsgebieten als Basis zum besseren Verständnis bereits anthropogen beeinflusster Gebiete dienen kann. Das hier untersuchte Gebiet hat daher das Potential zum Referenzgebiet für zukünftige Studien und Forschungsprojekte. Bedingt durch die Kürze der vorliegenden Abfluss- und Niederschlagszeitreihen war es nötig, den bestehenden Datenmangel durch eine Vielzahl von experimentellen Methoden und Ansätzen auszugleichen. Dieser Ansatz wird im Folgenden der "Multi-Methoden-Ansatz" genannt. Die ausgewählten Methoden sollten dabei so viele Aspekte der Abflussbildung abdecken wie möglich. Es wurden daher nicht nur Zeitreihen von Abfluss, Niederschlag, Bodenfeuchte- und Grundwasserdynamik gemessen, sondern auch eine große Zahl an Kurzzeitmessungen und Experimenten durchgeführt. Diese beinhalteten u.a. Messung des Bestandesniederschlags, Bestimmung der Wasserchemie, Bestimmung bodenphysikalischer Parameter und der Bodenmineralogie, sowie geophysikalische Messungen und Tracermethoden. Die Synthese der Resultate gleicht dem Zusammensetzen eines Puzzles. Das so entstandene Bild des dynamischen Prozess-Ensembles ist trotz möglicher fehlender Puzzlestücke hochinformativ. In einem nächsten Schritt wurden die ausgewählten Methoden im Hinblick auf Erkenntnisgewinn und Kosten (d.h. finanzielle Kosten und Arbeitszeit) evaluiert. Das durch die experimentellen Ergebnisse gewonnene Bild der Abflussbildung wurde anschließend mit Hilfe des physikalisch basierten Modells Catflow überprüft. Weiterhin wurde mit dem prozessbasierten Modell Wasim-ETH der Einfluss der Landnutzung auf die Abflussbildung auf Einzugsgebietsskala untersucht. Die Ergebnisse des "Multi-Methoden-Ansatzes" zeigen, dass die Abflussreaktion in diesem Gebiet sehr schnell erfolgt. Vertikales präferenzielles Fliessen ist hier von großer Bedeutung und wird in den Sommermonaten noch durch Hydrophobizitätseffekte verstärkt. Schneller lateraler Fluss im Untergrund ist eine weitere Vorraussetzung für die schnelle Reaktion des Abflusses (Oberflächenabfluss ist hier aufgrund der hohen hydraulischen Leitfähigkeiten unwahrscheinlich). Obwohl bei der Untersuchung der Bodenfeuchtedynamik in einigen Profilen laterale Fließmuster beobachtet wurden, ist die Art und Lage der Untergrundstrukturen, die auf der Hangskala schnellen lateralen Fluss verursachen, noch unklar und sollte genauer untersucht werden. Die Tatsache, dass bei Niederschlagsereignissen der Großteil der Niederschlagsmenge nicht zum Abfluss kommt (>90%, oft auch >95%), sowie der kontinuierliche Abfluss selbst nach Wochen der Trockenheit, lassen auf einen großen unterirdischen Speicher schließen. Der Wechsel von Winter (nass) zu Sommer (trocken) scheint Veränderungen im Prozess-Ensemble hervorzurufen, die sich in der Änderung von Fließmustern, von Grundwasser-Oberflächenwasser-Interaktionen, sowie veränderter Reaktion der Wasserchemie auf Niederschlagsereignisse beobachten ließ. Die Modellstudie bestätigte die Bedeutung der schnellen Fließwege. Als Folge von Informationsdefiziten über die Strukturen des Untergrunds ließ sich jedoch die Abflussbildung noch nicht vollständig reproduzieren. Die Untersuchung zur Bedeutung der Landnutzung für die Abflussbildung mit Hilfe eines Einzugsgebietsmodells zeigte die Zunahme der maximalen Abflüsse mit zunehmender Entwaldung. Weiterhin erwies sich auch die Lage der abgeholzten Flächen als ein wichtiger Faktor für die Abflussreaktion. Der "Multi-Methoden-Ansatz" lieferte wichtige Erkenntnisse zum Verständnis der Abflussbildungspozesse in den Anden Südchiles und zeigte sich als adäquates Mittel für hydrologische Prozess-Studien in datenarmen Gebieten.

No description available.

Pollution of surface and groundwater, due to improper land management, has become a major problem worldwide. Integrated watershed modelling provides a tool for the understanding of the processes governing water and matter transport at different scales within the watershed. The Soil Water Assessment Tool (SWAT) has been successfully utilized for the combined modelling of water fluxes and quality within a large range of scales and environmental conditions across the world. For suitable assessments integrated watershed models require large data sets of measured information for both model parameterization as for model calibration and validation. Data scarcity represents a serious limitation to the use of hydrologic models for supporting decision making processes, and may lead unsupported statements, poor statistics, misrepresentations, and, ultimately, to inappropriate measures for integrated water resources management efforts. In particular, the importance of spatially distributed soil information is often overlooked. In this thesis the eco-hydrological SWAT model was been applied to assess the water balance and diffuse pollution loadings of two rivers within a rural context at the mesoscale watershed level: 1) the Western Bug River, Ukraine, 2) the Águeda River, Portugal. Both watersheds in focus serve as examples for areas where the amount and quality of the measured data hinders a strait forward hydrologic modelling assessment. The Dobrotvir watershed (Western Bug River, Ukriane) is an example of such a region. In the former Soviet Union, soil classification primarily focused on soils of agricultural importance, whereas, forested, urban, industrial, and shallow soil territories were left underrepresented in the classification systems and resulting soil maps. Similarly the forest-dominated Águeda watershed in North-Central Portugal is a second example of a region with serious soil data availability limitations. Through the use of pedotransfer functions (PTFs) and the construction of soil-landscape models the data gaps could be successfully diminished, allowing a subsequent integrated watershed modelling approach. A valuable tool for the data gap closure was the fuzzy logic Soil Land Inference Model (SoLIM) which, combined with information from several soil surveys, was used to create improved maps. In the Dobrotvir watershed the fuzzy approach was used to close the gaps of the existing soil map, while in the Águeda watershed a new soil properties map, based upon the effective soil depths of the landscape, was constructed. While the water balance simulation in both study areas was successful, a calibration parameter ensemble approach was tested for the Águeda watershed. In the common modelling practice the individual best simulation and best parameter set is considered, the tested approach involved merging individual model outputs from numerous acceptable parameter sets, tackling the problematic of parameter equifinality. This procedure was tested for both original soil map and the newly derived soil map with differentiation of soil properties. It was noticeable that a better model set-up, with a better representation of the soil spatial distribution, was reflected in tighter model output spreads and narrower parameter distances. A further challenge was the calibration of water quality parameters, namely nitrate-N in the Dobrotvir watershed and sediment loads in the Águeda watershed. The limited amount of water quality observations were handled by assessing and by process verification at the smallest modelling unit, the hydrological response unit (HRU). The ruling hydrological processes could be depicted by combining own measured data and modelling outputs. The management scenario simulations showed the anticipated response to changes in management and reflected the rational spatial variation within the watershed reasonably well. The impacts of the different intervention options were evaluated on water balance, nitrate-N export and sediment yield at the watershed, sub-watershed and, when feasible, HRU level. This thesis covers two regional case studies with particular data limitations and specific processes of water and matter fluxes. Still, data reliability is a problem across the globe. This thesis demonstrates how relevant it is to tackle shortages of spatially differentiated soil information. The considered approaches contribute toward more reliable model predictions. Furthermore, the tested methods are transferable to other regions with differing landscape and climate conditions with similar problems of data scarcity, particularly soil spatially differentiated information.

<p>  </p> <p>The management of water resources in East Africa is inherently challenged by rainfall variability and the uneven spatial distribution of freshwater resources. In addition to these issues, meteorological and water data collection has been inconsistent over the past decades, and unclearly defined purposes or end goals for collected data have left many datasets ineffectively curated. In light of the data intensiveness of current modelling and planning methods, data scarcity and inaccessibility have become substantial impediments to informed decision-making. Among the outputs of this research are 1) a revised technique for evaluating bias correction performance on reanalysis data for use in regions where precipitation data is temporally discontinuous which can potentially be applied to other types of climate data as well, 2) a new methodology for quantifying qualitative information contained in legislation and official documents and websites for the assessment of relationships between documented meteorological and water data policies and resulting outcomes in terms of data availability and accessibility, and 3) a fresh look at data needs and the value data holds with respect to water resources decision-making and management in the region.</p>

Flooding is of great concern world-wide, causing damage to infrastructure, property and loss of life. Low-income countries, in particular, can be negatively affected by flood events due to their inherent vulnerabilities. Moreover, data to perform studies for flood risk management in low-income regions are often scarce or lacking sufficient quality. This thesis proposes new methodologies and explores the use of unconventional sources of information in flood hazard assessment in areas where the quantity or sufficient quality of traditional hydrometrical data are lacking.  One method was developed to account for errors in spatially averaged rainfall, from a sparse rain-gauge network, used as input to a rainfall-runoff model. A spatially-averaged and event-dependent rainfall depth multiplier led to improvements of the hydrographs at calibration. And by using a distribution of the multiplier, identified from previous events in the catchment, improvement in predictions could also be obtained. A second method explored the possibility of reproducing an unmeasured extreme flood event using a combination of models, post-event data, precipitation and an uncertainty-analysis framework. This combination allowed the identification of likelihood-associated parameter sets from which the flood hazard map for the extreme event could be obtained. A third and fourth study made at the regional scale explored the value of catchment similarities, and the effects of climate on the hydrological response of catchments. Flood frequency curves were estimated for 36 basins, assumed ungauged, using regional information of short flow records, and local information about the frequency of the storm. In the second regional study, hydro-climatic information provided great value to constrain predictions of series of daily flow from a hydrological model. Previously described methods, used in combination with unconventional information within an uncertainty analysis, proven to be useful for flood hazard assessment at basins with data limitations. The explored data included: post-event measurements of an extreme flood event, hydro-climate regional information and local precipitation data. The methods presented in this thesis are expected to support development of hydrological studies underpinning flood-risk reduction in data-poor areas. / Extremt höga vattenflöden ställer till stora problem i hela världen. De skadar infrastruktur och egendom och orsakar död. Framför allt kan låg- och medelinkomstländer vara väldigt sårbara för extrema flöden. I dessa länder saknas dessutom ofta data som behövs för att kunna bedöma översvämningsrisker, eller så finns bara data av dålig kvalitet. Denna avhandling föreslår nya metoder som använder okonventionella informationskällor vid bedömning av översvämningsrisker i områden där traditionella hydrologiska data saknas eller har otillräcklig kvalitet. En metod utvecklades för att ta hänsyn till fel i rumslig medelnederbörd beräknad från ett glest nät av nederbördsmätare att användas som indata i en nederbörds-avrinningsmodell. Användning av en multiplikator för medelvärdesbildad nederbörd, i tid och rum, för enskilda högflödestillfällen ledde till förbättrad modellkalibrering. Genom att använda multiplikatorfördelningar, identifierade från tidigare högflödestillfällen i avrinningsområdet, kunde också prognoser förbättras. En andra metod använde sig av möjligheten att reproducera ett extremt högflöde inom ramen för en osäkerhetsanalys med hjälp av en kombination av modeller, nederbördsdata och data som uppmätts i efterhand. Denna kombination gjorde det möjligt att identifiera parametervärdesuppsättningar med hophörande sannolikheter ur vilka det gick att erhålla en översvämningskarta för det höga flödet. En tredje och fjärde studie i regional skala utforskade värdet av likheter mellan avrinningsområden och hur områdenas hydrologiska gensvar beror av klimatet. Kurvan för kumulativa högflödesfrekvenser (flood frequency curve, FFC) kunde skattas med hjälp av lokal nederbördsinformation och regional information om korta tidsserier av vattenföring från 36 avrinningsområden som antogs sakna vattenföringsdata. I den andra regionala studien visade sig hydroklimatisk information av värde för att avgränsa godtagbara prognoser för daglig vattenföring från en hydrologisk modell. Tidigare beskrivna metoder, använda tillsammans med okonventionell information inom ramen för en osäkerhetsanalys, visade sig vara användbara för att bedöma översvämningsrisker i avrinningsområden med databegränsningar. Bland utforskade data fanns: mätningar i efterhand av ett extremt högflöde, hydroklimatisk regional information och lokala nederbördsmätningar. Metoderna i denna avhandling förväntas kunna stödja utvecklingen av hydrologiska studier av höga flöden och översvämningar i områden med bristande datatillgång. / Las inundaciones ocasionan daños a la infraestructura, propiedad y pérdida de vidas a nivel mundial. Los países en desarrollo son los más vulnerables a inundaciones, la calidad y cantidad de datos hidro-climatológicos disponibles en los mismos dificulta el desarrollo de estudios para la evaluación de riesgo a esta amenaza. Esta tesis propone métodos en la que se hace uso de fuentes de información no-convencionales para la evaluación de riesgo por inundación en regiones con datos escasos o limitados. Un método considera el error asociado a la precipitación promedio sobre cuencas en modelos lluvia-escorrentía como un factor multiplicador del histograma del evento. El uso de la precipitación promedio junto con una distribución probabilística del factor multiplicador como datos de entrada a un modelo de lluvia-escorrentía mejoraron los hidrogramas durante los periodos de calibración y predicción. Un segundo método exploró la posibilidad de reproducir un evento extremo de inundación usando una combinación de modelos hidrológicos e hidráulico, un análisis de incertidumbre, datos hidrométricos recopilados después del evento y datos de precipitación registrados durante-el-evento. Dicha combinación permitió la identificación de los parámetros de los modelos y la elaboración un mapa de amenaza por inundaciones para dicho evento. Adicionalmente, se estimaron curvas de frecuencia de inundaciones para 36 cuencas, asumidas no aforadas, mediante un método de regionalización que usa registros de caudal de corta duración disponibles en la región. Dichas curvas fueron extendidas haciendo uso de información local sobre la frecuencia de las tormentas. Se encontró que la información hidro-climatológica tiene un gran valor para reducir el rango de incertidumbre de las simulaciones de caudal diaria de un modelo hidrológico. Los métodos anteriores se usaron en combinación con información no-convencional dentro de un análisis de incertidumbre y han probado su utilidad para la evaluación de riesgo por inundaciones en cuencas con registros escasos o limitados. Los datos utilizados en esta tesis incluyen datos hidrométricos recopilados pasado el evento, registros hidro-climatológicos regionales y precipitación local. Se espera que los métodos presentados aquí contribuyan al desarrollo de estudios hidrológicos importantes para la reducción del riesgo por inundaciones en regiones con déficit de registros hidro-climatológicos.