Hybrid machine learning for flood prediction: comparing CHIRPS satellite and ground station data

Flooding in Jakarta is a multifaceted issue influenced by a combination of geographical, social, economic, and environmental factors. This study focuses on predicting floods by comparing automatic rain gauge (ARG) ground station data and Climate Hazards Group InfraRed Precipitation (CHIRPS) satellite data using the Adaptive Neurofuzzy Inference System (ANFIS) integrated with principal component analysis (PCA). The dataset includes precipitation measurements from both ARG and CHIRPS along with water level data spanning from 2014 to 2020. ARG provides precise local rainfall data, while CHIRPS offers extensive regional precipitation coverage. To enhance data quality, preprocessing techniques such as mean imputation, data normalisation, and the interquartile range (IQR) method were employed. The ANFIS-PCA model, which integrates fuzzy logic and neural network training, was applied using an 80:20 split for training and validation. When trained with ARG ground station data and water level measurements, the ANFIS-PCA model demonstrated superior accuracy, achieving a root mean square error (RMSE) of 0.13, mean absolute error (MAE) of 0.12, and R² of 0.82. In contrast, the ANFIS model without PCA yielded higher errors, with RMSE 6.3, MAE 6.2, and R² 0.74. Training with CHIRPS satellite data resulted in significantly higher errors (RMSE 30.14, MAE 24.05, R² 0.42). These findings underscore the superiority of ground-based measurements for flood prediction, given the reduced precision and higher susceptibility to errors in satellite-derived data. While CHIRPS satellite data offers broader spatial coverage, its limitation in precision and higher susceptibility to errors reduce its effectiveness for accurate flood prediction.