Optimasi Estimasi Data Curah Hujan Berbasis Satelit terhadap Data Pengamatan Hujan menggunakan Random Forest di Wilayah Balikpapan

Abstract

Unstable precipitation estimation from Satellite Precipitation Products poses a challenge for local hydrological analysis. This study evaluates the accuracy of three raw satellite products known as GPM IMERG, CHIRPS, and GSMaP across daily, ten day (dasarian), and monthly scales against ground data from Synoptic, AWS, and ARG stations in Balikpapan. To improve data precision, a machine learning based bias correction utilizing the Random forest algorithm was implemented. The baseline evaluation revealed notable systematic errors, marked by low to moderate linear relationships on raw scatterplot diagrams. Raw GPM IMERG consistently overestimated rainfall, showing a positive percentage bias of 74.77 percent on the monthly scale at the ARG Kariangau. Conversely, raw CHIRPS and GSMaP showed underestimation trends, with monthly negative biases of 19.25 percent and 23.65 percent at the Synoptic station. These discrepancies caused the initial Nash Sutcliffe Efficiency values to remain low or negative. The application of the Random forest algorithm significantly enhanced quantitative accuracy. Post correction time series curves aligned closely with field observations, and scatterplot data converged along the ideal 1:1 line. The error distribution curves narrowed sharply around zero millimeters, proving that error variance was minimized. Furthermore, the model successfully lowered the False alarm Ratio close to zero for moderate to heavy daily rainfall events. Among the uncorrected datasets, raw GSMaP provided the most stable baseline performance. After calibration, the RF GSMaP variant maintained its superiority as the most accurate model across all scales. Its optimal performance occurred on the monthly scale at the Synoptic station, achieving a high correlation coefficient r of 0.96, a low Mean Absolute Error of 19.92 mm, a Root Mean Square Error of 35.66 mm, a minimized bias of 0.39 percent, and an exceptional Nash Sutcliffe Efficiency of 0.91. These results demonstrate that the Random forest framework is highly effective for calibrating satellite rainfall data over complex local regions