Rainfall prediction for climate-resilient agriculture: a robust ensemble with SARIMA and LightGBM

Agriculture is a vast area with many different practices and requirements based on the crop, region, climate, and other aspects. The inputs like seeds, fertilizer, pesticides, and environmental conditions like climate, soil quality, and water availability strongly influence agriculture. When it pertains to rainfall, many crops depend on the quantity and timing of precipitation. Typically, the growth of plants, the germination of seeds, and the production of fruits and grains all require a certain amount of rainfall. However, issues like water logging, erosion, drought, and crop failure can also result from either too much or too little rainfall. In this study, we investigate the significance of rainfall prediction and propose an ensemble methodology for improving rainfall forecast accuracy for the benefit of paddy cultivation in Tripura. The proposed model involves utilizing the Seasonal Autoregressive Integrated Moving Average model ensembled with Light Gradient Boosting Machine (LightGBM). The model is developed and trained using more than 20 years of relevant weather data. We also compared the model with other state-of-the-art models that are already in use. The study demonstrated the effectiveness of SARIMA and LightGBM models both in stand-alone and ensembled modes. The accuracy improvement of 8% and the R2-score of 83% in the results impressively show the usefulness of the ensemble. Additionally, the proposed model's improved performance in terms of root mean square error and mean absolute error, with lower values of 72.09 and 55.242 mm, respectively, supports its feasibility and appropriateness for long-term rainfall forecasts.