Sequential Hybrid Approach for Reliable Detection of Rainfall Pauses at the Beginning of the Rainy Season in Senegal: Towards a Predictive Tool for False Starts

In semi-arid regions, particularly in the Sahel, agricultural drought—defined as a water deficit affecting crops and their productivity—constitutes a major risk to food security [1]. This region, and notably Senegal, is characterized by strong interannual rainfall variability [2], frequent delays in the onset of the rainy season, an extended dry season, and a continuous increase inland surface temperature (LST) [3, 4]. These climatic changes disrupt traditional agricultural cycles and increase the vulnerability of rainfed production systems.

A critical yet often overlooked phenomenon in this context is the false onset of the rainy season. It occurs as a series of weak initial rains—sometimes totaling about 20 mm over 2–3 days—followed by a dry spell. This sequence can create a false sense of security for farmers, who may sow prematurely, exposing their crops to early-season water stress that can significantly reduce yields. Despite extensive research on drought detection, the false onset—a key factor in shifting the rainfall calendar— has received relatively little attention, limiting the predictive capabilities of existing models [5]. Early detection is therefore essential to adjust sowing schedules, mitigate water stress risks, and support farmers’ decision-making. Recent advances in remote sensing and machine learning have greatly enhanced drought monitoring and forecasting. Remote sensing provides continuous spatio-temporal information despite the scarcity of in situ measurements [4]. Vegetation and soil moisture indices such as NDVI, VHI, LSWI, and SIF have been widely used to assess vegetation water stress [6–9]. Moreover, composite indices integrating precipitation, temperature, and remote sensing data (e.g., CDI [10], SMADI [11], IDSI [12]) offer improved insights into drought dynamics.

In parallel, machine learning models have shown strong potential for drought and rainfall forecasting, particularly when combined with physical constraints that enhance predictive realism [9, 13]. Explainable Artificial Intelligence (XAI) techniques further highlight the relative importance of climatic variables, showing for example that temperature can be more influential than precipitation [1].

Within the Sahelian context, the timing of the rainy season is strongly influenced by large-scale climate phenomena such as ENSO. Some approaches, such as Kohonen maps, have been used to classify the onset and cessation of rainy seasons [5], while logistic regression models based on climate predictors (SST, precipitable water, dew point temperature, winds) have been proposed for seasonal onset forecasting [14]. However, these statistical models often show limited interregional transferability and struggle to capture complex intra-seasonal dynamics—particularly those linked to false starts.

To address these challenges, recurrent neural networks (RNNs) have been increasingly applied in hydrological and climatic studies. LSTM networks have proven effective for long-term temporal dependencies [15, 16], whereas GRU architectures better handle short-term fluctuations [20]. Transformer models, on the other hand, excel at capturing non-local relationships across long sequences. Hybrid architectures that integrate these models have recently emerged as a promising direction, leveraging heterogeneous data such as meteorological time series, soil properties, and remote sensing observations [18]. However, in West African contexts, the scarcity and heterogeneity of ground and satellite data remain a key limitation. Existing studies often overlook the false onset phenomenon, focusing instead on general drought indices or seasonal rainfall patterns. Statistical models alone lack the capacity for robust interregional generalization, while purely neural approaches may suffer from overfitting or reduced interpretability.

To overcome these limitations, this study proposes a hybrid neural architecture combining LSTM, GRU, and Transformer components for the detection and early prediction of false rainy season onsets in Senegal. The approach leverages multisource climatic and phenological data, enriched by statistical change-point detection tests (Pettitt, Kendall, Lombard) to identify structural shifts in rainfall, soil moisture, and vegetation dynamics. By integrating climatic, phenological, and statistical criteria, we define a False Onset Index (FOI) that enables more robust and realistic detection of false starts, improving the reliability of seasonal forecasts and supporting farmers in optimizing their sowing calendars.

The remainder of this paper is structured as follows: Section 2 describes the dataset, the LSTM–GRU–Transformer hybrid model, and hyperparameter optimization. Section 3 presents the results, including performance evaluation, robustness analysis, SHAP-based interpretability, and ablation studies. Section 4 also discusses the implications, limitations, and perspectives of the study, and Section 5 concludes by summarizing the main contributions.