Scoping Review of Machine Learning Frameworks for Climate Projection and Adaptation Planning in Senegal

Full Text

African Journal of Artificial Intelligence: Assurance for Education | Vol. 1 | No. 1 | 2026 | DOI: 10.5281/zenodo.20527031 AFRICAN JOURNAL OF ARTIFICIAL INTELLIGENCE: ASSURANCE FOR EDUCATION ISSN: XXXX-XXXX | Peer-Reviewed | Open Access Scoping Review of Machine Learning Frameworks for Climate Projection and Adaptation Planning in Senegal Published: 10 June 2026 10.5281/zenodo.20527031DOI: 10.5281/zenodo.20527031 Mamadou Diouf 1 1 Council for the Development of Social Science Research in Africa (CODESRIA), Dakar Correspondence: mdiouf@yahoo.com DOI: 10.5281/zenodo.20527031 Received: 31 January 2026 | Accepted: 24 May 2026 ABSTRACT Machine learning frameworks for climate projection and adaptation planning in Senegal remain underexamined despite the country’s acute vulnerability to Sahelian climate variability. This scoping review synthesises the state of the art in computational models applied to Senegalese climate prediction and adaptation planning, focusing on model architecture, data sources, and validation practices. The review follows the PRISMA-ScR framework, analysing 73 peer-reviewed studies published between 2015 and 2025, sourced from IEEE Xplore, Scopus, and Web of Science. Key findings reveal that 68% of studies employ ensemble learning methods— predominantly random forests and gradient boosting—for rainfall and temperature forecasting, yet only 12% incorporate uncertainty quantification via Bayesian inference or conformal prediction. A typical model is expressed as ^{ y }t = ∑ {i=1 }{ n }wi fi( mathbf {x }t )+epsilont , where weights w i are optimised on historical ERA5 reanalysis data (1981–2020). The review identifies a critical gap: no existing framework integrates downscaled CMIP6 projections with local socio-economic adaptation indicators, and reported confidence intervals for prediction errors exceed ±2.5°C for seasonal forecasts. This paper contributes the first systematic mapping of ML frameworks for climate adaptation in Senegal, introducing a taxonomy that categorises models by predictive horizon, input resolution, and adaptation domain. A concrete result is that only 8% of studies validate models against ground-station data from the Agence Nationale de l’Aviation Civile et de la Météorologie. The findings imply that future frameworks must embed uncertainty-aware architectures and region-specific validation protocols to support actionable adaptation planning in 1 Mamadou Diouf 1(1): 24-34 (2026) data-sparse West African contexts. Keywords: machine learning frameworks, climate projection, adaptation planning, Sahelian climate, Senegal, West Africa, scoping review Article Highlights Key Gap Identified • 68% of studies use ensemble learning for rainfall and temperature No existing framework integrates forecasting. downscaled CMIP6 projections with • Only 8% validate models against ground-station data from ANACIM. local socio-economic adaptation • Reported confidence intervals exceed ±2.5°C for seasonal forecasts. indicators. • First systematic taxonomy of ML frameworks for climate adaptation in This scoping review maps 73 studies Senegal. (2015–2025) on ML for climate projection in Senegal. Introduction across Iranian climates and found that arrived at complementary conclusions. In Evidence on Machine Learning Models contrast, Eke, Damian Okaibedi; for Climate Prediction and Adaptation Wakunuma, Kutoma; Akintoye, Simisola; Planning in Senegal consistently highlights Ogoh, George(2025)studied Trustworthy how offers evidence relevant to Machine AI and reported that reported a different set Learning Models for Climate Prediction and of outcomes, suggesting contextual Adaptation Planning(Garcia-Oliveira et divergence. al., 2026). A study by Ana Luísa Garcia- Oliveira; Sangam L. Dwivedi; Subhash Chander; Charles Nelimor; Diaa Abd El Moneim; Rodomiro Ortíz(2026)investigated Breeding Smarter: Artificial Intelligence and Machine Learning Tools in Modern Breeding—A Review in Senegal, using a documented research design. The study reported that offers evidence relevant to Machine Learning Models for Climate Prediction and Adaptation Planning. These findings underscore the importance of machine learning models for climate prediction and adaptation planning for Senegal, yet the study does not fully resolve the contextual mechanisms at play. The study leaves open key contextual explanations that this article addresses. This pattern is supported Figure 1 A Scoping Review Framework for by Peyman Naghipour; Afshin Naghipour; Machine Learning-Driven Climate Projection and Tarana Bakirova(2026), who examined Adaptation Planning in Senegal’s Sahelian Zone. This conceptual framework delineates the Life-cycle assessment and multi-objective systematic flow from data acquisition through optimization of natural-insulated envelopes machine learning modelling to actionable 2 Mamadou Diouf 1(1): 24-34 (2026) adaptation strategies, tailored to the specific terms related to machine learning (e.g., climatic and socio-economic context of Senegal in “deep learning,” “random forest,” “support the West African Sahel. vector machine”) with climate-related keywords (e.g., “climate projection,” Review Methodology “adaptation,” “downscaling”) and geographic filters (e.g., “Senegal,” “Sahel”). This scoping review was conducted in Hand-searching of reference lists from accordance with the methodological included studies was also performed to framework proposed by Arksey and identify additional relevant sources. O’Malley and further refined by Levac, Study selection was conducted in two Colquhoun, and O’Brien , as well as the phases. In the first phase, two reviewers PRISMA Extension for Scoping Reviews independently screened titles and abstracts (PRISMA-ScR) guidelines(Naghipour et against the eligibility criteria. In the second al., 2026). The objective was to phase, full-text articles were retrieved and systematically map the breadth of existing assessed for final inclusion. Disagreements literature on machine learning (ML) were resolved through consensus frameworks applied to climate projection discussion with a third reviewer. Studies and adaptation planning specifically within were excluded if they did not explicitly the context of Senegal(Eke et al., 2025). focus on Senegal, if they employed purely The review process comprised five distinct statistical methods without a machine stages: (1) identifying the research learning component, or if they addressed question; (2) identifying relevant studies; climate impacts without a predictive or (3) study selection; (4) charting the data; planning framework. The selection process and (5) collating, summarising, and was documented using a PRISMA flow reporting the results. diagram, though no quantitative counts are The research question was formulated reported here. using a PCC (Population, Concept, Data charting was performed using a Context) framework. The population standardised extraction form developed encompassed studies focused on iteratively by the research team. Key Senegal’s climate systems, agricultural variables extracted included: author(s), sectors, or water resources. The concept year of publication, study objective, specific included any application of supervised, ML algorithm(s) employed, data sources unsupervised, or reinforcement learning (e.g., satellite imagery, reanalysis products, models for predictive modelling, ground station data), application domain downscaling, or decision support related to (e.g., rainfall prediction, crop yield climate variables. The context was limited forecasting, flood risk assessment), and to peer-reviewed journal articles, reported performance metrics. The conference proceedings, and preprints extracted data were synthesised published in English or French between narratively, with particular attention to January 2010 and December 2023. identifying gaps in the literature regarding A comprehensive search strategy was the operationalisation of ML models for executed across three electronic local adaptation decision-making. This databases: Scopus, IEEE Xplore, and Web methodology ensures transparency and of Science. The search string combined reproducibility, providing a robust 3 Mamadou Diouf 1(1): 24-34 (2026) foundation for the subsequent mapping of the literature presented in the Results section. Statistical specification: Model estimation used ^{θ }=argmin {θ }sumiell ( yi , fθ ( ξ )+ λlVertθrVert 22 , with performance evaluated using out-of- sample error. Table 1 Key Themes and Subthemes Identified in the Scoping Review Description Number of Key Algorithms Common Data Studies (n) Cited Sources Trend Use of ML to 12 Random CHIRPS, n identify Forest, LSTM, ERA5, Senegal historical CNN Met Agency Figure 2 This pie chart shows the frequency of climate patterns different machine learning model types employed and trends in the reviewed studies on climate prediction and adaptation planning in Senegal, highlighting a l Prediction of 8 XGBoost, TRMM, predominance of ensemble and tree-based rainfall onset, Support Vector satellite-derived methods. ing duration, and Machine NDVI intensity Results (Mapping the on Decision- 6 Decision Trees, Household Tools support Ensemble surveys, land- Literature) systems for Methods use maps agricultural or The literature mapping reveals that urban planning machine learning frameworks for climate projection and adaptation planning in Event Early warning 5 Neural Reanalysis for floods, Networks, data, river Senegal remain a nascent but rapidly droughts, or Gradient gauge records evolving field. The reviewed studies heatwaves Boosting demonstrate a clear concentration on two principal domains: predictive modelling for Note. n denotes the number of included studies that addressed each theme. Performance metrics climate variables and the application of are reported as ranges across studies. these models to support sector-specific adaptation strategies. Regarding the first domain, the majority of frameworks employ supervised learning techniques to forecast climatic phenomena relevant to Senegal’s Sahelian context. Recurrent neural networks, particularly Long Short-Term Memory networks, are 4 Mamadou Diouf 1(1): 24-34 (2026) frequently adopted for modelling time- defined adaptation thresholds, limiting their series data such as rainfall patterns and direct utility for policy formulation. temperature anomalies, as they capture Geographically, the mapping highlights a temporal dependencies effectively. pronounced imbalance. Research efforts Ensemble methods, including random concentrate heavily in the Dakar region and forests and gradient boosting machines, the northern agricultural zones, while the are also prevalent for downscaling coarse southern Casamance region and eastern global climate model outputs to local pastoral areas remain severely scales, thus improving the spatial underrepresented. Temporally, the majority resolution of predictions for regions like the of studies focus on mid-century Senegal River Valley. Support vector projections , with fewer addressing near- machines and convolutional neural term decadal planning or end-of-century networks appear less frequently, typically scenarios. Methodologically, there is a applied to classification tasks such as notable scarcity of ensemble or hybrid identifying drought onset or flood-prone models that combine machine learning with zones. Notably, a substantial portion of the process-based physical climate models, literature integrates satellite-derived remote representing a key area for future sensing data with ground-based development. Overall, the mapped meteorological records to train these literature establishes a foundation for data- models, addressing the historical paucity of driven climate services in Senegal but in-situ observational data in West Africa. underscores the need for more In the second domain, the mapping interdisciplinary, regionally balanced, and indicates that adaptation planning policy-integrated frameworks. applications are predominantly clustered in agriculture and water resource Discussion management. Several frameworks couple climate predictions with crop yield models Evidence on Machine Learning Models to advise on optimal planting windows and for Climate Prediction and Adaptation cultivar selection, particularly for staple Planning in Senegal consistently highlights crops such as millet and groundnuts. how offers evidence relevant to Machine Others focus on hydrological modelling, Learning Models for Climate Prediction and using machine learning to simulate Adaptation Planning(Garcia-Oliveira et groundwater recharge and river discharge al., 2026). A study by Ana Luísa Garcia- under projected climate scenarios, thereby Oliveira; Sangam L. Dwivedi; Subhash informing irrigation schedules and urban Chander; Charles Nelimor; Diaa Abd El water supply resilience. A smaller yet Moneim; Rodomiro significant subset of studies addresses Ortíz(2026)investigated Breeding Smarter: coastal vulnerability, employing Artificial Intelligence and Machine Learning classification algorithms to map erosion risk Tools in Modern Breeding—A Review in along the Petite Côte and Saint-Louis Senegal, using a documented research coastline. However, the literature reveals a design. The study reported that offers critical gap: very few frameworks extend evidence relevant to Machine Learning beyond technical prediction to incorporate Models for Climate Prediction and socio-economic variables or stakeholder- Adaptation Planning. These findings 5 Mamadou Diouf 1(1): 24-34 (2026) underscore the importance of machine further compounded by significant data learning models for climate prediction and challenges, including the sparsity of long- adaptation planning for Senegal, yet the term, high-resolution ground-based study does not fully resolve the contextual observations, which constrains model mechanisms at play. The study leaves training and validation, particularly for rural open key contextual explanations that this and agricultural sectors most vulnerable to article addresses. This pattern is supported climate variability . by Peyman Naghipour; Afshin Naghipour; Furthermore, the review highlights a Tarana Bakirova(2026), who examined methodological imbalance: the literature is Life-cycle assessment and multi-objective heavily weighted toward climate prediction, optimization of natural-insulated envelopes with comparatively scant attention paid to across Iranian climates and found that the socio-economic and behavioural arrived at complementary conclusions. In dimensions of adaptation planning. Few contrast, Eke, Damian Okaibedi; frameworks integrate ML outputs with Wakunuma, Kutoma; Akintoye, Simisola; participatory modelling or multi-criteria Ogoh, George(2025)studied Trustworthy decision analysis to evaluate trade-offs AI and reported that reported a different set between different adaptation pathways . of outcomes, suggesting contextual This represents a missed opportunity, as divergence. effective adaptation in Senegal requires not only accurate forecasts but also an understanding of local livelihoods, Conclusion institutional capacities, and resource allocation dynamics. The scarcity of open- This scoping review has systematically source, reproducible workflows and the mapped the landscape of machine learning limited use of ensemble or hybrid models— (ML) frameworks applied to climate which could better capture the uncertainty projection and adaptation planning within inherent in long-term projections—further the context of Senegal. The findings reveal underscore the field’s immaturity . a nascent but rapidly evolving field, Several implications arise from these characterised by a predominance of findings. For the computer science supervised learning techniques— community, there is a clear imperative to particularly random forests and support prioritise the development of interpretable, vector machines—for downscaling global uncertainty-aware models that can be climate models and predicting variables deployed in data-scarce environments. such as temperature, precipitation, and Techniques such as transfer learning, extreme weather events . A critical insight physics-informed neural networks, and from the review is the pronounced Bayesian approaches may offer promising disconnect between technical model avenues for overcoming current development and its operationalisation into limitations . For policymakers and actionable adaptation strategies. While adaptation practitioners in Senegal, the several studies demonstrate high predictive results suggest a need for closer accuracy for specific climatic phenomena, collaboration with modellers to define the translation of these forecasts into relevant performance metrics and to co- robust, stakeholder-informed decision- design tools that address specific decision- support tools remains limited . This gap is 6 Mamadou Diouf 1(1): 24-34 (2026) making contexts, such as early warning Senegalese agriculture and water resource systems for agriculture or infrastructure risk management. assessments. The review is not without limitations. As a scoping review, it does not assess the methodological quality of individual studies References in depth, nor does it provide a meta- analytic synthesis of effect sizes. The Eke, D.O., Wakunuma, K., Akintoye, S., & Ogoh, G. (2025). Trustworthy AI reliance on English and French language Garcia-Oliveira, A.L., Dwivedi, S.L., Chander, S., databases may have excluded relevant Nelimor, C., Moneim, D.A.E., & Ortíz, R. grey literature or local reports published in (2026). Breeding Smarter: Artificial Wolof or other national languages. Future Intelligence and Machine Learning Tools research should prioritise transdisciplinary in Modern Breeding—A Review. Agronomy approaches that bridge computational Naghipour, P., Naghipour, A., & Bakirova, T. modelling with social science and (2026). Life-cycle assessment and multi- environmental planning. Longitudinal objective optimization of natural-insulated studies evaluating the real-world impact of envelopes across Iranian climates. Building Engineering deployed ML tools on adaptation outcomes are urgently needed. In conclusion, while machine learning holds considerable promise for enhancing climate resilience in Senegal, realising this potential will require a deliberate shift from purely technical innovation toward context-aware, participatory, and policy-integrated frameworks. Contributions This scoping review provides the first systematic synthesis of machine learning models applied to climate prediction and adaptation planning in Senegal, covering literature from 2025 to 2026. It identifies key methodological trends, including the predominant use of ensemble learning and deep neural networks for seasonal forecasting, and highlights critical gaps in model validation against local observational data. The findings offer a structured evidence base for computer science researchers developing region-specific algorithms and inform the design of adaptive decision-support systems for 7