In Ghana, flash floods are often triggered by severe storms. Flood Early Warning Systems (FEWS) can mitigate flood impacts but require accurate, near real-time rainfall data. Implementing FEWS in Ghana is challenging due to sparse ground-based data and a lack of accurate, rainfall data with short latency. While satellite-based rainfall products offer a promising alternative, they often show significant discrepancies compared to ground observations, limiting their use for effective FEWS.

The Meteosat satellite images provide a valuable source of data for near-real-time applications, due to its short latency (within 15 minutes) and relatively high temporal (15-minute) and spatial resolution of (3 x 3 km at sub-satellite point). This study explores the use of Earthformer, a space-time transformer model, to improve rainfall intensity estimates with minimal latency, using data from the Meteosat satellite. By evaluating the potential of Earthformer to create rainfall estimates, this research aims to contribute to more reliable FEWS, ultimately strengthening disaster risk management in flood-prone regions of
Ghana.

The Earthformer model is trained on IMERG-Final, a satellite-based rainfall product known for its relative high accuracy but with delayed availability of several months. With the application of FEWS in mind, it was investigated whether the model could be adapted to improve the estimation of higher rainfall intensities. Two model setups were tested: one using a mean squared error (MSE) loss function during the training of the model and another with a balanced weighting loss function to emphasize higher intensities.

The model’s accuracy was first evaluated by comparing its outputs with IMERG-Final on a test dataset, and secondly with ground station observations from the Trans-African Hydro-Meteorological Observa-tory (TAHMO), and Ghana Meteorological Services (GMET) for the year 2022. IMERG-Early was used as a benchmark for near real-time performance. The comparison with IMERG-Final revealed that both Earthformer models outperformed IMERG-Early for lower rainfall intensities in terms of probability of detection (POD), success rate (SUCR) and the Critical Succes Index (CSI) and that the balanced loss model also outperformed IMERG-Early for higher intensities. However, as rainfall intensity increased, the performance of both Earthformer models and IMERG-Early decreased.

Further comparisons with ground station data highlighted weak correlations at 30-minute intervals between all satellite rainfall estimates and ground observations (including IMERG-Final, IMERG-Early and both Earthformer models). However, when the data was aggregated to daily intervals, correlations improved significantly, suggesting that timing errors could play a role, however, further investigation is needed to quantify their impact.

Additional analysis revealed that peak rainfall was often underestimated, while lower intensities were overestimated. This discrepancy could be explained by several factors: the coarser spatial resolution of satellite estimates compared to gauge stations, the displacement of rainfall from observed clouds, and difficulty in capturing warm rain processes. Additionally, it was concluded that capturing spatial variability within Mesoscale Convective Systems (MCSs) is challenging. This is potentially due to anvil cloud tops obstructing the satellite’s view, similar cloud-top temperatures for different rainfall intensities, and strong wind shear increasing the risk of rainfall misallocation.

This research demonstrates the potential of a space-time transformer model for near real-time rainfall estimation as it shows improved performance for most intensities when compared to IMERG-Early, with IMERG-Final set as the reference truth. However, the reduced performance at higher intensities and discrepancies with ground observations of all satellite based products underscore the need for further model development to improve extreme rainfall detection and better align satellite estimates with ground truth data. These improvements are essential for the model’s utility in FEWS and contributing to more
effective disaster risk management in flood-prone regions.

This research addresses how Ricanau Mofo, a low-lying village in Surinam, can become a more water-adaptive and sustainable village, while it faces land erosion, river bank erosion, changing rainfall patterns and sea level rise. It is urgent to intervene, as these issues are expected to increase in occurrence due to climate change. Constraints and limitations that are important to take into account are cultural preservation, maintaining accessibility to the Cottica River, the limited availability of financial resources and the need for a low-maintenance intervention.

Three strategies are proposed. The first focuses on addressing land erosion. Planting vegetation on critically eroding areas is a short term measure, while the long-term involves constructing footpaths with drainage channels. This not only mitigates soil erosion, but also regulates water and is relatively cost efficient. The second strategy targets the river bank erosion, which includes wooden bulkheads with vegetation and stones for short term implementation. For the long term, a river bank protection system with groynes is designed, to break waves, slow down the stream velocity and in time causes land gain. As the long term plans require external financial aid, a business case is set up and shared with the captain of Ricanau Mofo, STEORR, the District Commissioner and the Ministry of Public Works. The third strategy addresses water damage in the urban environment. Short-term it consists of providing building guidelines of where to build more water adaptive, and how. This is placed on an informational board in the village. The long-term contains a flood early warning system and recommended equipment.

This project's significance lies in identifying interventions that are effective against erosion and water damage while being locally implementable in the rural areas of Surinam. It can be seen as a pilot project that is scalable to other villages along the Cottica River or in the whole of Surinam. However, there are limitations to the project. The most important is the lack of data quantity and data quality. This caused implications for dimensioning the interventions and their financial impact. Another limitation is that the project does not create ‘dry feet’ for the village; it creates a way of mitigating water damage while living next to the Cottica River. In addition, there is a limitation in the financing of follow-up projects. Therefore, a business case is also being delivered to the Ministry of Public Works, the District Commissioner of Marowijne South-West and the captain of Ricanau Mofo. They can use it to apply for funds from international organisations and include it in future policy plans.

To summarise, Ricanau Mofo can become more water adaptive by regulating how to build and where, by continuing the prototype of the bulkheads by planting more vegetation and, by requesting financial aid for the river bank protection long term. As Figure 0.2 shows, it not only contributes as a report, but also in a tangible form of a prototype and in educational information boards in the local language to enhance the continuity and help the village.