Hydrological connectivity describes the internal linkages between runoff generation in upper parts of the catchment and the receiving water. It is quantified as the ratio of the runoff reaching the catchment s outlet and the total internal runoff generation. It thus effectively bridges t he gap between the point-scale separation of rainfall into soil water st orage and (sub-)surface runoff as opposed to what we see as response at the hillslope or catchment scale. In between both, significant water red istribution from runoff source areas into sinks may occur such that isol ated active areas in upslope regions may become disconnected and may not contribute to the actual outflow. This pattern-process interaction is o ne of the main reasons why hydrological observations at laboratory or pl ot scale are inadequate to explain the phenomena witnessed on hillslopes and in catchments and is why hydrological connectivity has become, in t he last decade, a central concept in hydrology, particularly in semi-ari d environments. The overall purpose of this dissertation is to identify which factors control the connectivity of a hillslope and to d iscuss how this reflects on data collection and modelling.While traditionally hydrological variables are estimated through point observa tions dynamically varying in time, connectivity stresses the need to als o account for space. It is the spatiotemporal heterogeneity in infiltrat ion capacity and rainfall intensity that determines connectivity or disc onnectivity with complex and spatially varying thresholds governing whic h parts of the domain are active and which contributing and which domina nt processes affect storage, redistribution and connectivity. Once estab lished that it is the spatial pattern of heterogeneous runoff generation and abstraction and not merely the statistical distribution of this het erogeneity, we ask ourselves which aspects of pattern determine the conn ectivity and which are negligible. We test a number of hypotheses by opp osing landscape metrics with the modelled hydrograph of a virtual hillsl ope. Each metric describes a different characteristic of the pattern and their variable correlation with connectivity can thus supply an objecti ve criterion to ascertain what controls connectivity. The aggregation of runoff source areas in interaction with the flow distance to the outlet dominate in the presence of a mosaic of runoff sinks and sources. When the area reacts more homogeneous, e.g. due to high rainfall intensity, t he pattern disappears and it becomes mainly the travel time distribution that governs connectivity. While the used landscape metrics could only partially predict the connectivity, they are insightful tools for hypoth esis testing and, we believe, can reversely be used in basin classificat ion and the identification of dominant processes. If we establish a link between a certain metric and a particular basin functioning, we may als o decide upon a range of values of that metric that a basin should fulfil in order to be classified in a particular class.We propose ran domness of pattern as an important characteristic of heterogeneity to di stinguish two classes: when the heterogeneity is random, its particular configuration becomes superfluous and only the statistical distribution of its properties remains of interest. If, on the other hand, the hetero geneity expresses clear spatial configuration or gradients, accounting f or this configuration is indispensible. This has important implications for modelling as well as for data collection. If the particular spatial configuration has no influence on the outcome, it also does not need to be explicitly modelled. The degrees of freedom of the model can therefor e be reduced to a smaller number of parameters that set the statistical distribution rather than having one uncorrelated parameter for each spat ial element in the modelling space which would clearly lead to the probl em of equifinality. If we can only parameterise those spatial aspects th at matter, we effectively reduce the overparamaterisation that plagues m odern distributed modelling. As for data collection: since the exact spa tial configuration of random patterns is redundant knowledge, a sparse d ata collection that settles the statistical distribution of heterogeneit y suffices in that case; while in the presence of significant configurat ion, data sampling strategies should be tailored to capture the spatial patterns.If we want to gain a better understanding of proces ses, we need to build models that can test hypotheses and collect data t hat allows to falsify the hypotheses. Both the reconfirmation of space a s a crucial dimension beside time and the observation that non-random pa ttern matters, emphasises the need in connectivity related problems for data that can sample dense spatial variations in hydrological state vari ables and processes. We explore a combination of Electrical Resistivity Tomography (ERT) and Time Domain Reflectometry (TDR): ERT samples spatia l averages and can be translated to a dense 3D resistivity distribution. TDR, on the other hand, is a point-scale measurement and is used to sup ply a transformation between resistivity and water content, the so-calle d pedophysical relation. By collecting data both before and after a conc entrated flow experiment in a semi-arid gully, we map the total infiltra tion that occurred by taking the difference between the estimated water distributions before and after. The straightforward three-step scheme (E RT inversion, transformation to water content, subtraction) returns, how ever, an invalid solution. To improve the outcome, we developed an alter native inversion that constrains the outcome to comply to our knowledge of the experiment and that jointly inverses ERT and TDR while optimising the pedophysical parameters. Although these alterations effect signific ant improvements, the signal-to-noise ration in the data and the poor re solution of the outcome undermine the reliability of the obtained infiltration map. We were therefore unable to utilise it in further hypothesis testing or modelling. Nonetheless, we believe that ERT and TDR sampling in combination with our proposed inversion scheme can potentially give insightful results. nrpages: 140 status: published