Data Volume as the Primary Constraint in Tropical Rainfall Forecasting: A Comparative Analysis

Abstract

North Sumatra’s tropical climate requires accurate daily rainfall forecasting for paddy cultivation. This study evaluates three deep learning architectures — Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and CNN-BiLSTM — alongside XGBoost for daily rainfall forecasting in Deli Serdang Regency, North Sumatra, Indonesia. Two datasets were used: a two-year observational dataset from the Indonesian Climatological Station or known as BMKG (788 samples, January 2024 – March 2026) and an eleven-year NASA POWER reanalysis dataset (4,082 samples, January 2015 – March 2026), both at coordinates 3.6211°N, 98.7149°E. All models were evaluated using RMSE, MAE, MSE, and R² under a chronological 70/15/15% train-validation-test split. On BMKG dataset, all models achieved severely limited performance (highest R² = 0.0774, XGBoost), When trained on the larger NASA POWER dataset, all models exhibited consistent RMSE reductions of 63.89%-66.71%, with XGBoost achieving the best overall performance (RMSE = 7.5461 mm, R² = 0.1504) and the only positive R² across both datasets. Among deep learning architectures, BiLSTM consistently yielded the best R². The persistently low R² across all models reflects the fundamental challenge posed by zero-inflated, discontinuous nature of tropical rainfall distributions. A Wilcoxon signed-rank test confirmed that the performance differences between datasets were statistically significant for all models at α = 0.05 (p = 0.0020–0.0371), with XGBoost and CNN-BiLSTM R² remaining significant after Bonferroni correction. These findings recommend BiLSTM as the promising deep learning candidate for future investigation with extended observational records.

GS Cited Analysis