Small to medium-sized man-made freshwater reservoir are a reliable source for drinking water supply, hydropower generation and irrigation purposes worldwide. However, water volumes in these reservoirs can be significantly affected by prolonged droughts, resulting in severe impacts on society (Kozacek, 2014; Mahr, 2018). To mitigate the impact of such events it is crucial for decision makers to know when the available water resources are lacking. Although many reservoirs are closely monitored, this data is not always readily available. Inadequate information sharing, inaccessibility, and a lack of tools to predict future reservoir storages contribute to this problem.

Remote sensing has the potential to address this problem. The Global Water Watch is a platform that and provides earth-observed surface area dynamics that can be used to monitor small to medium-sized reservoirs worldwide and detect trends in water availability. While this method serves as a valuable indicator of water availability, it falls short in providing decision-makers with the necessary absolute volume time series and volume predictions. Currently, no platform exists beyond in-situ measurements to meet this essential need.

This thesis presents a novel method for retrieving near real-time volume time series in small to medium-sized man-made reservoirs worldwide using remotely sensed open data. The method utilises the MERIT-Hydro digital elevation model, HydroMT and stream flow methods by Eilander et al. (2023), and literature by Messager et al. (2016) to reconstruct reservoir bathymetry. This novel approach in reconstructing reservoir bathymetry enables the conversion of available reservoir area time series into volume time series. These were employed in autoregressive and multi-linear regression models to predict water availability up to six months in advance. The models incorporate ERA5 precipitation data by Hersbach's (2020) and the Standardised Precipitation and Evaporation Index (SPEI) by Beguería et al. (2021) to improve the accuracy of the volume predictions.

When comparing the novel method to the method proposed by Messager et al. (2016), the novel method yielded more accurate reservoir volume estimations. The method successfully obtained bathymetries and accurate volume estimations when validating using 2 reservoirs in Zambia and 48 in India, demonstrating the potential of this novel approach. However, some reservoirs with complex shapes faced initial delineation challenges, resulting in inaccurate volume predictions. These issues could be resolved by manually delineating the area for bathymetry reconstruction. Moreover, regression models were applied to case study reservoirs in Eswatini and Lesotho, demonstrating reasonable predictive capabilities with the Heidke Skill Scores ranging from 0.77 to 1 for up to 2 months ahead. However, precise prediction of extreme decreases in reservoir levels requires a physically based approach that incorporates the volumetric time series provided by this novel method. The study emphasises the necessity of considering the volume time series’ memory to predict water availability and provides a valuable foundation for volume time series analysis using remotely sensed data.

Anticipatory action requires models that can accurately and reliably predict the impact of natural hazards. However, impact forecasts are often underestimated when consecutive hazards are not considered. In the Bicol region in the Philippines, typhoons trigger 90% of landslides, causing a lot of fatalities and damage to infrastructure and agriculture. The lack of information on past landslide events has hampered the construction of landslide forecasting models. Currently, a machine learning (ML) impact-based forecasting (IBF) model for typhoons is operational in the Philippines. The model was developed by 510, an initiative of the Netherlands Red Cross. The model predicts impact due to the high wind speeds associated with typhoons and includes the possible impact due to landslides only via a static landslide susceptibility map. Hence, this study focused on extending the 510 typhoon model via hybrid modeling into a multi-hazard forecasting model for both typhoons and landslides to improve the forecast by considering impact from typhoon-induced landslides. The implementation of the hybrid multi-hazard impact-based forecasting model was tested on two typhoon events in the Bicol region.

A hydrometeorological landslide IBF model was successfully created, even with the limited data on landslide occurrences and rainfall available. The newly established regional event duration threshold for Bicol was applied on the case study events with an increased impact boundary of 300 km compared to the typhoon impact boundary of 100 km. The results of the hybrid multi-hazard model showed an improved impact forecast -compared to the model considering solely static input of landslides, which underestimated impact- in both location extent of the impact forecast and in accuracy: the True Positives doubled, whereas the False Negatives reduced by half. The separate landslide IBF model as an extension of the existing ML typhoon model provided additional benefits as these models can be decoupled to optimize the performance and reliability of both. This study resulted in the prototype of an impact-based multi-hazard model for typhoons and landslides for the Philippines and demonstrated the importance of considering impact from consecutive hazards.

Most people of Malawi are dependent on rainfed agriculture for their livelihoods. This leaves them vulnerable to drought and changing rainfall patterns due to climate change. Over time, farmers have adopted local strategies and knowledge that help reducing the overall vulnerability to climate variability shocks. One other option to increase the resilience of rainfed farmers to drought, is providing forecast information on the upcoming rainfall season. Forecast information has the potential to inform farmers in their decisions surrounding agricultural strategies. However, significant challenges remain in the provision of forecast information. Often, the forecast information is not tailored to farmers, resulting in limited uptake of forecast information into their agricultural decision-making. Therefore, this study explores whether drought forecast information can be linked to existing farmers strategies and local knowledge on predicting future rainfall patterns. During a period of three months in Malawi, participatory research approaches are used to create an understanding of what requirements drought forecast information should meet to effectively inform farmers in their decision-making. Consequently, a sequential threshold model was established that relates annually monitored meteorological indicators before the rainy season, to the occurrence of dry conditions during the season. Dry conditions were expressed in the drought indicators that farmers require for their agricultural decision-making. Additionally, using interviews among stakeholders and a visualisation of the current information flow, further insights on the current drought information system were developed. Although farmers have their own strategies and timing of decision-making, this research has generalized some of the opinions and strategies to develop the ‘requirements’ which a contextualized forecast should meet. In August farmers require a prediction of the onset of the rainy season, typically starting mid-November. In addition, an update on the timing of the onset of rains is required in beginning of November. An overall indication of the ‘dryness’ of the rainy season is required in September. Here, ‘dryness’ is characterized by the number of dry spells, a composite ‘drought index’ of associated rainfall variables by the farmers. The forecast should be on a scale that is locally relevant (EPA level). This research consequently established a forecasting model, based on meteorological variables from local knowledge which can complement the forecast variables from the DCCMS. The results of forecast verification show that meteorological indicators based on local knowledge have a predictive value for forecasting drought indicators. Subsequently, skill analysis of forecasting incorporating all the above dimensions shows that the accuracy of the forecast differs per location with an increased skill to the Southern locations. In addition, it is also location dependent whether the contribution of wind, temperature or ENSO indicators gives the most predictive value. The results show that a combination of all indicators have the best predictive value. In addition, the results show that local knowledge indicators have an increased predictive value in forecasting the locally relevant critical events in comparison to the currently used ENSO-related indicators by the DCCMS. Additional research is needed to further analyse certain aspects of this research, such as research on the robustness of the model used. Research on the risk farmers are willing to take in their respective decisions could act as another requirement the forecast skill should meet. This highlights the importance of having continuous feedback from the farmers, since farmers may experience adverse impacts from wrongly informed decisions. Despite these limitations, it is argued that the inclusion of local knowledge in the current drought information system of Malawi may improve the provision of forecast information for farmers and shows that it is possible to capture local knowledge in a technical approach. The findings have relevant implications for other stakeholders, such as humanitarian and meteorological organisations, that are implementing drought-risk reduction approaches and climate services.