Geographia Technica, Vol 21(2), Special Issue: Artificial Intelligence Applications in Geography, 2026, pp. 97-118
PERFORMANCE OF MACHINE LEARNING AND DEEP LEARNING MODELS FOR PREDICTING RAINFALL IN A LARGE WATERSHED: CASE STUDY OF BENGAWAN SOLO RIVER BASIN, INDONESIA
Jumadi JUMADI 
, Kuswaji Dwi PRIYONO , Ali Hasan ABDULLAH, Supari SUPARI , Hamza AIT ZAMZAMI , Farha SATTAR , Muhammad NAWAZ , Hamzah HASYIM , Steve CARVER
ABSTRACT: Predicting rainfall in large, heterogeneous watersheds remains among the most important hydrological challenges. This research investigates the effectiveness of both ML (Machine Learning) and DL (Deep Learning) for predicting spatiotemporal rainfall in the Bengawan Solo watershed, Indonesia. Satellite rainfall data from CHIRPS (spatial resolution: 0.05°) were prepared and sampled for the period 1981–2024. The data set contained 523 grid points. We employed nine ML and DL algorithms: Random Forest (RF), Extreme Gradient Boosting (XGB), Support Vector Regression (SVR), Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), Convolutional Neural Network (CNN), and Transformer. Models were trained on the samples from 1981 to 2019 and tested on 2020–2024. Performance was judged from mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). XGB showed the best overall performance (MAE ≈59 mm; R² ≈0.73). GRU became the most competitive DL model at 60 mm (MAE ≈60 mm; R² ≈0.72). Temporal model analysis shows that XGB and GRU stay among the top three models with the minimum monthly errors. TCN, CNN, and Transformer exhibited higher errors and more monthly variability. XGB and GRU have average MAE values of ~59–60 mm and R² values of ~0.71–0.72 across most grids. MAE values for TCN, CNN, and Transformer are greater than 76 mm, while R² values are lower. The data we obtained indicate that using ensemble decision tree models and recurrent neural networks across large tropical areas yields greater stability and more reliable spatiotemporal rainfall predictions than more sophisticated DL architectures.
Keywords: Rainfall prediction; Machine learning; Deep learning; CHIRPS; Bengawan Solo; Spatial analysis; Temporal analysis.
Full article here