<![CDATA[Rainfall variability has a major impact on the agricultural sector because conditions that are too wet or too dry can increase the risk of flooding, drought, crop disturbances, and changes in the planting calendar. Subseasonal to Seasonal (S2S) forecasting is important because it is located between short-term weather forecasts and seasonal predictions, so that it can provide early information for decision-making in the agriculture, water management, and disaster mitigation sectors. This study aims to develop a deep learning-based S2S rainfall prediction model using the CNN–LSTM and CNN–GRU hybrid architecture by utilizing ECMWF atmospheric variables, as well as CHIRPS rainfall as predictors. CNNs are used to extract spatial features, while LSTM and GRU model temporal dynamics. The model generates predictions at a lead time of 0–45 or up to 46 days ahead for each ECMWF release. The results of the evaluation showed spatial variation in performance on the island of Java. CNN–LSTM showed the best performance, especially in the central Java region with an RMSE of 3.286, a correlation of 0.944, a BSS of 0.414, and a CRPSS of 0.775. CNN-GRU also showed good performance in the central to southern regions with an RMSE of 2.939, a correlation of 0.954, a BSS of 0.406, and a CRPSS of 0.631. In general, CNN-LSTM provides a more stable performance, especially in probabilistic evaluations, while the western Java region still shows greater prediction challenges.]]>
Copyrights © 2026
