Eugene C. Ezin

This study addresses the challenges of end-to-end (E2E) Speech-to-Text Translation (STT) for the low-resource Fongbe-to-French language pair using a transfer learning approach. We first establish robust baselines by integrating state-of-the-art pretrained speech encoders (HuBERT-147, AfriHuBERT, XLS-R, Whisper) with powerful text decoders (mBART, NLLB). This initial phase identified the AfriHuBERT–NLLB and XLS-R–NLLB combinations as the most competitive E2E configurations. To further enhance performance, we propose and evaluate three hybrid feature fusion strategies, the Bidirectional Co-Attention (BCOAT), the Feature-wise Linear Modulation (FiLM), and the Feature Sum (SUM). These methods are designed to strategically fuse intermediate representations extracted from two distinct and powerful encoders, AfriHuBERT and XLS-R, within the E2E architecture. The fusion process aims to enrich the linguistic and tonal information critical for accurate translation of the tonal Fongbe language. Experimental results demonstrate significant performance gains over the baselines. The BLEU score improved from 26.32 to a peak of 27.78 for the AfriHuBERT–NLLB configuration (using FiLM), and from 26.27 to a maximum of 28.05 for the XLS-R–NLLB configuration (using SUM). These findings confirm that translation quality for tonal languages like Fongbe can be substantially improved by extracting and combining high-quality, complementary features through advanced encoder fusion. Our hybrid feature fusion methods present a substantial advance in speech translation quality within resource-scarce linguistic environments.

Background: Food insecurity affects a large proportion of the world's population, with enormous disparities to the disadvantage of low-income countries, especially those in sub-Saharan Africa. In this context, public decision-makers, funders, and NGOs urgently need streamlined indicators on households in precarious situations to better orient their policies and aid. However, existing food security monitoring tools focus on isolated dimensions of the problem, and some remain static. We aim to fill these gaps by focusing on a composite metric of food security to explore associations and identify variables with predictive relevance for food security outcomes using machine learning techniques. Methodology: This paper applies machine learning methods to a panel dataset from Burkina Faso to explore which agroecological practices are associated with improvements in food security and have an impact on the prediction of food security outcomes. Results: Overall, the trained models recorded accuracies varying from 64% for the Multi-Layer Perceptron model to 72% for the assembly models, with learning times generally < 1 min, except for the SVM model, which recorded around 30 min, suggesting good potential for these models to identify food-insecure households. Our results show that agroecological intensification through a combination of multiple agroecological practices, such as livestock and crop species diversification, plays a key role in predicting households’ food security status. Indeed, households practicing agroecology with good species diversification recorded a food insecurity rate of 6.6%, compared with 15% for conventional households, and pairwise differences in the prevalence of food security are significant (p-values most often < 0.01). The partial dependence plots show that the probability of belonging to the food-secure class increases by 3% as species richness increases or as livestock are integrated. Conclusion: The performance of the algorithms using a simplified measure of food security suggests that machine learning model-driven approaches could significantly improve food insecurity crisis response.

With the impacts of climate change, floods have become increasingly frequent in recent years. Estimating flood hazard thresholds and peak floodwater levels based on flood frequency analysis is crucial for anticipating and preparing for potential flooding events. This study aims to estimate flood hazard thresholds, flood occurrence probabilities, and the return periods of peak floodwater levels in the Nokoue lake watershed in Benin. To achieve this, the standardized water level index, also known as the Flood hazard Index, was calculated to estimate flood hazard thresholds. The three best probability distribution models, Gumbel, Generalized Extreme Value (GEV), and Generalized Pareto (GPA), were selected to project future floodwater levels using annual maximum daily water level data for extreme floods from 1997 to 2022, obtained from a water gauge site at Nokoue lake. Three goodness-of-fit tests were applied to identify the best-fitting probability distribution model: a Taylor diagram (three-dimensional analysis), a cumulative probability density diagram based on the root-mean-square error (RMSE), and an L-moment diagram (two-dimensional analysis). The Flood hazard Index values ranged from −1.10 to +3.40, with 77.78% showing positive indices and 22.22% showing negative indices. The flood hazard thresholds were classified in ascending order of index values: limited hazards, moderate hazards, significant hazards, and critical hazards. The analysis results indicate that the flood hazard thresholds are defined as follows: below 3.94 m for limited hazards, from 3.94 m up to 4.04 m for moderate hazards, from 4.04 m to 4.14 m for significant hazards, and above 4.14 m for critical hazards. The distribution model analysis showed that the Gumbel distribution best fits the Nokoue lake watershed, with an RMSE of 0.0724, compared to 0.0754 and 0.0761 for the GEV and GPA models, respectively. The annual maximum daily water levels for various non-exhaustive return periods, 2, 3, 5, 10, 25, 50, and 100 years, were estimated and compared. The return period for the highest recorded annual maximum daily water levels (4.4 m/day) in the Nokoue lake watershed were calculated to be 12, 15, and 15 years using the Gumbel, GEV, and GPA models, respectively. Quantile analysis revealed that the Gumbel distribution produced overestimated results compared to the GEV and GPA models for return periods exceeding 10 years. Exceptional and very exceptional hydrological events have return periods of 100 and 150 years, corresponding to peak flow levels of 4.95 m and 5.05 m respectively. Finally, the results of this study will be invaluable for flood hazard managers in monitoring flood alerts and for water resource engineers in determining dimensions for designing flood control structures such as spillways, dams, and bridges, thereby improving the management of recurrent flooding events.

A flood is defined as a relatively high flow rate or water level in a river or body of water, characterized by an elevation above the norm, leading to the inundation of low-lying areas or bodies of water that are not normally submerged. Almost every year, floods of significant magnitude impact the Nokoue lake basin, endangering nearby populations as they greatly affect the economy and infrastructure development. The unresolved issue of flooding in Nokoue lake necessitates a thorough study and analysis of specific flow models, particularly those concerning the frequency, probability, and scale of flooding. In this context, the present article on threshold determination and flood frequency analysis is highly valuable in mitigating the chronic flooding problem of Nokoue lake. Flood frequency analysis involves interpreting the historical flood data and predicting future probabilities. Annual data on the maximum water level of Nokoue lake over 27 years, from 1997 to 2022, was used for this purpose. The peak water level indices of Nokoue lake and classification were used to characterize flood thresholds. The Gumbel extreme value distribution method, the generalized extreme value distribution method, and the generalized Pareto method were used to test the probability of flood occurrence. The analysis results indicate that flood hazard thresholds are defined as follows: from negative infinity up to below 3.94 m for limited risk, from 3.94 m up to 4.04 m for moderate risk, from above 4.04 m to 4.14 m for significant risk, and from above 4.14 m to 4.95 m for critical risk. Among the flood frequency analysis methods, the Gumbel extreme value distribution is considered more effective and suitable for estimating flood occurrence probabilities and for designing flood control measures for Nokoue lake with a root mean square error (RMSE) of 0.0724, compared to 0.0754 and 0.0761 for the GEV and GPA distributions, respectively. The position and scale parameters (φ; α) of the Gumbel distribution were estimated to be 3.802 and 0.249, respectively. This allows for the calculation of the probability of an extreme water level occurring within a return period in a single day. Thus, the extreme water levels (flood quantiles) associated with return periods of 10, 50, and 100 years, as determined by the Gumbel distribution, are 4.36m, 4.77m, and 4.95m, respectively. Exceptional and very exceptional hydrological events have return periods of 100 and 150 years, corresponding to peak flow levels ranging from 4,95m and 5,05m respectively. These values are of crucial importance for the design of flood prevention structures (infrastructure) intended to mitigate flood risk.

The evolution and acceleration of the effects of climate change on the water cycle demand adaptation and mitigation plans. Consequently, the implementation of public policies (management, governance, and strategy) regarding flood risk management or prevention (such as levee heights and spillway dimensions of dams) is based on the characterization of flood hazards. This requires understanding extreme climate phenomena and conducting a rigorous probabilistic analysis of hydrometric data. This study aims to estimate the flood quantiles for lake Nokoue. To achieve this, the adopted methodology involved fitting the Generalized Extreme Value (GEV) distribution, the Gumbel distribution, and the Generalized Pareto (GPA) distribution to the annual maximum water levels of lake Nokoue from 2015 to 2022. We estimated experimental probabilities using the Weibull formula. The assessment of the fit quality of the theoretical probability distributions to our data sample indicated that the Gumbel distribution was the most suitable, with a root mean square error (RMSE) of 0.0724, compared to 0.0754 and 0.0761 for the GEV and GPA distributions, respectively. The position and scale parameters (φ; α) of the Gumbel distribution were estimated to be 3.802 and 0.249, respectively. This allows for the calculation of the probability of an extreme water level occurring within a return period in a single day. Thus, the extreme water levels (flood quantiles) associated with return periods of 10, 50, and 100 years, as determined by the Gumbel distribution, are 4.36m, 4.77m, and 4.95m, respectively. These values are of crucial importance for the design of flood prevention structures (infrastructure) intended to mitigate flood risk.