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

Abstract— This work focuses on the detection of false onsets of the rainy sea son in Senegal, a critical factor that can lead farmers, particularly smallholders, to initiate agricultural activities prematurely. Such errors, caused by misleading early rainfall events, result in yield losses and increase farmers’ vulnerability to climate variability. Unlike existing methods, our approach incorporates statistical tests (such as Pettitt, Kendall, and Lombard) to enrich the input dataset with relevant change points related to rainfall, soil moisture, and vegetation. This enrichment step, combined with a formal detection of false onsets based on climatic, phenological, and statistical criteria, enhances the relevance, robustness, and contextualization of detection compared to purely statistical or physical approaches. In this context, a deep learning methodology was developed to identify false onsets at an early stage using multivariate climatic data. We designed a hybrid model combining LSTM, GRU, and multi-head attention layers to extract complementary representations of the input sequence. Model hyperparameters were optimized through Bayesian search to enhance detection performance. Results show consistent improvements across all key metrics: accuracy increased from 0.84 to 0.88, F1-score from 0.833 to 0.86, recall remained perfect at 1.0, precision rose from 0.767 to 0.81, and AUC improved from 0.900 to 0.92. These gains demonstrate the overall robustness of the optimized model, ensuring more reliable detection of false onsets.